isolated_alloc: directly use mmap for allocations

[nia-e]

In case there isn't a resolution to #4357, this will fix it at least. Benchmark with the allocator force-enabled vs baseline:

Comparison with baseline (relative speed, lower is better for the new results):
  backtraces: 1.10 ± 0.06
  big-allocs: 0.94 ± 0.07
  mse: 0.96 ± 0.03
  range-iteration: 0.99 ± 0.05
  serde1: 1.03 ± 0.05
  serde2: 1.02 ± 0.05
  slice-chunked: 0.99 ± 0.04
  slice-get-unchecked: 0.93 ± 0.04
  string-replace: 0.84 ± 0.04
  unicode: 1.03 ± 0.04
  zip-equal: 0.95 ± 0.07

[RalfJung]

This is odd, since the resulting Layout is not what we are actually using -- we can't even tell mmap an alignment.

So really the "number of pages to allocate" is the only thing this should return I think.

[nia-e]

Made it return the size directly instead of page count since that's what we use, but good point. Also got rid of the page_layout function since we only care about the size when mmapping