Hint compiler to not specialize compile methods

[mofeing]

compile methods generate the same return types no matter the arguments. By adding @nospecialize and maybe @nospecializeinfer, maybe we can reduce the TTFX on compile when changing functions.

[github-actions]

Reactant.jl Benchmarks

Benchmark suite	Current: 222625f	Previous: 4f4cb40	Ratio
ViT base (256 x 256 x 3 x 32)/forward/CUDA/Reactant	1318940641 ns	1427801876 ns	0.92
ViT base (256 x 256 x 3 x 32)/forward/CUDA/Lux	232704274 ns	241937991 ns	0.96
ViT base (256 x 256 x 3 x 32)/forward/CPU/Reactant	5297983590 ns	5706546145 ns	0.93
ViT base (256 x 256 x 3 x 32)/forward/CPU/Lux	27391225052 ns	19592540088 ns	1.40
ViT small (256 x 256 x 3 x 4)/forward/CUDA/Reactant	1291926315.5 ns	1367040220.5 ns	0.95
ViT small (256 x 256 x 3 x 4)/forward/CUDA/Lux	8101212 ns	8765018 ns	0.92
ViT small (256 x 256 x 3 x 4)/forward/CPU/Reactant	1625149502 ns	1617045921.5 ns	1.01
ViT small (256 x 256 x 3 x 4)/forward/CPU/Lux	2291206144 ns	2228339749.5 ns	1.03
ViT tiny (256 x 256 x 3 x 32)/forward/CUDA/Reactant	1334446605.5 ns	1369778487.5 ns	0.97
ViT tiny (256 x 256 x 3 x 32)/forward/CUDA/Lux	93860961 ns	95384074 ns	0.98
ViT tiny (256 x 256 x 3 x 32)/forward/CPU/Reactant	2169485862 ns	2267863268 ns	0.96
ViT tiny (256 x 256 x 3 x 32)/forward/CPU/Lux	4271876468 ns	4660275687.5 ns	0.92
ViT tiny (256 x 256 x 3 x 4)/forward/CUDA/Reactant	1316968758 ns	1374725656.5 ns	0.96
ViT tiny (256 x 256 x 3 x 4)/forward/CUDA/Lux	7464719.5 ns	7919007 ns	0.94
ViT tiny (256 x 256 x 3 x 4)/forward/CPU/Reactant	1471509052 ns	1523146334.5 ns	0.97
ViT tiny (256 x 256 x 3 x 4)/forward/CPU/Lux	1559028201 ns	1657446563 ns	0.94
ViT tiny (256 x 256 x 3 x 16)/forward/CUDA/Reactant	1345744439.5 ns	1293396969 ns	1.04
ViT tiny (256 x 256 x 3 x 16)/forward/CUDA/Lux	11653564 ns	84573801 ns	0.14
ViT tiny (256 x 256 x 3 x 16)/forward/CPU/Reactant	1754843139 ns	1829891224 ns	0.96
ViT tiny (256 x 256 x 3 x 16)/forward/CPU/Lux	2487275309.5 ns	3023920644 ns	0.82
ViT small (256 x 256 x 3 x 16)/forward/CUDA/Reactant	1545602240 ns	1292433719 ns	1.20
ViT small (256 x 256 x 3 x 16)/forward/CUDA/Lux	95539755.5 ns	88498026 ns	1.08
ViT small (256 x 256 x 3 x 16)/forward/CPU/Reactant	2229052978 ns	2315496105 ns	0.96
ViT small (256 x 256 x 3 x 16)/forward/CPU/Lux	3716833384.5 ns	4181219607 ns	0.89
ViT small (256 x 256 x 3 x 32)/forward/CUDA/Reactant	1317185684 ns	1367617841 ns	0.96
ViT small (256 x 256 x 3 x 32)/forward/CUDA/Lux	122600223 ns	113824695 ns	1.08
ViT small (256 x 256 x 3 x 32)/forward/CPU/Reactant	3091138651 ns	3118475424 ns	0.99
ViT small (256 x 256 x 3 x 32)/forward/CPU/Lux	6307106053.5 ns	6707702872 ns	0.94
ViT base (256 x 256 x 3 x 16)/forward/CUDA/Reactant	1314407457 ns	1342010577 ns	0.98
ViT base (256 x 256 x 3 x 16)/forward/CUDA/Lux	128864521.5 ns	129882165.5 ns	0.99
ViT base (256 x 256 x 3 x 16)/forward/CPU/Reactant	3260848845 ns	3304575961 ns	0.99
ViT base (256 x 256 x 3 x 16)/forward/CPU/Lux	7279675017 ns	9641559937 ns	0.76
ViT base (256 x 256 x 3 x 4)/forward/CUDA/Reactant	1382526551 ns	1318534127.5 ns	1.05
ViT base (256 x 256 x 3 x 4)/forward/CUDA/Lux	86162139 ns	87600441 ns	0.98
ViT base (256 x 256 x 3 x 4)/forward/CPU/Reactant	2330506538 ns	1971771485 ns	1.18
ViT base (256 x 256 x 3 x 4)/forward/CPU/Lux	2552811939.5 ns	2525837503 ns	1.01

This comment was automatically generated by workflow using github-action-benchmark.

[github-actions]

Benchmark Results

	main	222625f...	main/222625f4c2fcab...
comptime/basics/2D sum	26.9 ± 2.1 ms	26.8 ± 1.1 ms	1
comptime/basics/Basic grad cos	0.0449 ± 0.0023 s	0.0447 ± 0.0015 s	1
comptime/basics/cos.(x)	31.2 ± 1.3 ms	0.0332 ± 0.0014 s	0.941
comptime/lux neural networks/ViT base	5.69 s	6.1 s	0.931
comptime/lux neural networks/ViT tiny	5.82 s	5.88 s	0.99
comptime/lux neural networks/vgg11 bn=false	0.389 ± 0.041 s	0.427 ± 0.019 s	0.911
comptime/lux neural networks/vgg13 bn=false	0.436 ± 0.02 s	0.43 ± 0.018 s	1.01
comptime/lux neural networks/vgg16 bn=false	0.521 ± 0.0026 s	0.515 ± 0.0041 s	1.01
comptime/lux neural networks/vgg19 bn=false	0.615 ± 0.03 s	0.612 ± 0.024 s	1
runtime/lux neural networks/ViT base (compiled)	6.4 s	6.45 s	0.993
runtime/lux neural networks/ViT tiny (compiled)	1.69 s	1.72 s	0.979
runtime/lux neural networks/vgg11 bn=false (compiled)	2.12 s	2.09 s	1.02
runtime/lux neural networks/vgg13 bn=false (compiled)	2.97 s	2.93 s	1.01
runtime/lux neural networks/vgg16 bn=false (compiled)	3.73 s	3.73 s	1
runtime/lux neural networks/vgg19 bn=false (compiled)	4.53 s	4.59 s	0.988
time_to_load	1.45 ± 0.021 s	1.45 ± 0.014 s	1

Benchmark Plots

A plot of the benchmark results have been uploaded as an artifact to the workflow run for this PR.
Go to "Actions"->"Benchmark a pull request"->[the most recent run]->"Artifacts" (at the bottom).