marginal improvement: atomic byte-counts

Eliminate an extra lock/unlock on the eviction path by making the
`nbytes` field atomic.

This produces some marginal performance improvements.
e.g. here are the benchmark results from my laptop:

```
[atomic_bytecount] % go test -benchtime 30s -bench . .
goos: linux
goarch: amd64
pkg: github.com/vimeo/galaxycache
cpu: Intel(R) Core(TM) i7-7600U CPU @ 2.80GHz
BenchmarkGetsSerialOneKey-4                             19998571              1849 ns/op            1536 B/op         28 allocs/op
BenchmarkGetsSerialManyKeys-4                           10693790              3401 ns/op            2394 B/op         46 allocs/op
BenchmarkGetsParallelManyKeys-4                         12335263              2872 ns/op            2381 B/op         45 allocs/op
BenchmarkGetsParallelManyKeysWithGoroutines/1-4         12733614              2967 ns/op            2382 B/op         45 allocs/op
BenchmarkGetsParallelManyKeysWithGoroutines/2-4         12099810              2755 ns/op            2379 B/op         45 allocs/op
BenchmarkGetsParallelManyKeysWithGoroutines/4-4         13770105              2689 ns/op            2378 B/op         45 allocs/op
BenchmarkGetsParallelManyKeysWithGoroutines/8-4         13695392              2717 ns/op            2374 B/op         45 allocs/op
BenchmarkGetsParallelManyKeysWithGoroutines/16-4                12850902              2745 ns/op            2367 B/op         45 allocs/op
BenchmarkGetsParallelManyKeysWithGoroutines/32-4                13667977              2749 ns/op            2356 B/op         45 allocs/op
BenchmarkGetsParallelManyKeysWithGoroutines/64-4                13518266              2917 ns/op            2330 B/op         44 allocs/op
PASS
ok      github.com/vimeo/galaxycache    398.172s
go test -benchtime 30s -bench . .  1201.40s user 37.27s system 309% cpu 6:39.64 total
                                                                                                                                                                                                                                                                                                  2022-08-03 11:55:31 ($? = 0)
[master] % go test -benchtime 30s -bench . .
goos: linux
goarch: amd64
pkg: github.com/vimeo/galaxycache
cpu: Intel(R) Core(TM) i7-7600U CPU @ 2.80GHz
BenchmarkGetsSerialOneKey-4                             20515902              1852 ns/op            1536 B/op         28 allocs/op
BenchmarkGetsSerialManyKeys-4                           10534585              3448 ns/op            2395 B/op         46 allocs/op
BenchmarkGetsParallelManyKeys-4                         11663724              2968 ns/op            2382 B/op         45 allocs/op
BenchmarkGetsParallelManyKeysWithGoroutines/1-4         12227192              2956 ns/op            2380 B/op         45 allocs/op
BenchmarkGetsParallelManyKeysWithGoroutines/2-4         12722972              2972 ns/op            2378 B/op         45 allocs/op
BenchmarkGetsParallelManyKeysWithGoroutines/4-4         10123365              3382 ns/op            2379 B/op         45 allocs/op
BenchmarkGetsParallelManyKeysWithGoroutines/8-4         11134406              3433 ns/op            2375 B/op         45 allocs/op
BenchmarkGetsParallelManyKeysWithGoroutines/16-4                12949555              3012 ns/op            2367 B/op         45 allocs/op
BenchmarkGetsParallelManyKeysWithGoroutines/32-4                12999852              2946 ns/op            2351 B/op         45 allocs/op
BenchmarkGetsParallelManyKeysWithGoroutines/64-4                13319569              3028 ns/op            2325 B/op         44 allocs/op
PASS
ok      github.com/vimeo/galaxycache    406.806s
go test -benchtime 30s -bench . .  1166.79s user 37.50s system 295% cpu 6:47.19 total
```

galaxycache.go

@@ -704,7 +704,7 @@ func (c *cache) stats() CacheStats {
 	c.mu.Lock()
 	defer c.mu.Unlock()
 	return CacheStats{
-		Bytes:     c.nbytes,
+		Bytes:     c.nbytes.Get(),
 		Items:     c.itemsLocked(),
 		Gets:      c.nget,
 		Hits:      c.nhit,
@@ -732,7 +732,7 @@ func (c *cache) add(key string, value valWithStat) {
 	c.mu.Lock()
 	defer c.mu.Unlock()
 	c.lru.Add(key, value)
-	c.nbytes += int64(len(key)) + value.size()
+	c.nbytes.Add(int64(len(key)) + value.size())
 }
 
 func (c *cache) removeOldest() {
@@ -745,9 +745,7 @@ func (c *cache) removeOldest() {
 }
 
 func (c *cache) bytes() int64 {
-	c.mu.Lock()
-	defer c.mu.Unlock()
-	return c.nbytes
+	return c.nbytes.Get()
 }
 
 func (c *cache) items() int64 {

galaxycache_test.go

@@ -388,8 +388,8 @@ func TestNoDedup(t *testing.T) {
 	testKStats := keyStats{dQPS: windowedAvgQPS{}}
 	testvws := g.newValWithStat([]byte(testval), &testKStats)
 	wantBytes := int64(len(testkey)) + testvws.size()
-	if g.mainCache.nbytes != wantBytes {
-		t.Errorf("cache has %d bytes, want %d", g.mainCache.nbytes, wantBytes)
+	if g.mainCache.nbytes.Get() != wantBytes {
+		t.Errorf("cache has %d bytes, want %d", g.mainCache.nbytes.Get(), wantBytes)
 	}
 }
 

typed_caches.go

@@ -53,7 +53,7 @@ func (c *candidateCache) get(key string) (*keyStats, bool) {
 type cache struct {
 	mu         sync.Mutex
 	lru        *lru.TypedCache[string, valWithStat]
-	nbytes     int64 // of all keys and values
+	nbytes     AtomicInt // of all keys and values
 	nhit, nget int64
 	nevict     int64 // number of evictions
 	ctype      CacheType
@@ -69,7 +69,7 @@ func newCache(kind CacheType) cache {
 func (c *cache) setLRUOnEvicted(f func(key string, kStats *keyStats)) {
 	c.lru.OnEvicted = func(key string, value valWithStat) {
 		val := value
-		c.nbytes -= int64(len(key)) + val.size()
+		c.nbytes.Add(-(int64(len(key)) + val.size()))
 		c.nevict++
 		if f != nil {
 			f(key, val.stats)

untyped_caches.go

@@ -57,7 +57,7 @@ func (c *candidateCache) get(key string) (*keyStats, bool) {
 type cache struct {
 	mu         sync.Mutex
 	lru        *lru.Cache
-	nbytes     int64 // of all keys and values
+	nbytes     AtomicInt // of all keys and values
 	nhit, nget int64
 	nevict     int64 // number of evictions
 	ctype      CacheType
@@ -73,7 +73,7 @@ func newCache(kind CacheType) cache {
 func (c *cache) setLRUOnEvicted(f func(key string, kStats *keyStats)) {
 	c.lru.OnEvicted = func(key lru.Key, value interface{}) {
 		val := value.(valWithStat)
-		c.nbytes -= int64(len(key.(string))) + val.size()
+		c.nbytes.Add(-(int64(len(key.(string))) + val.size()))
 		c.nevict++
 		if f != nil {
 			f(key.(string), val.stats)