Add SnoopPrecompile

[mcabbott]

Like https://github.com/FluxML/Zygote.jl/pull/1281 this adds https://github.com/timholy/SnoopCompile.jl

Timing some gradient of a model, it's not so obvious this saves time.

M1 mac, Julia 1.9- nightly

### Zygote v0.6.49, Flux v0.13.6

% time julia -e 'using Flux; gradient(m -> Flux.mse(m(rand(Float32,10,10,1,1)), rand(2,1)), Chain(Conv((3,3),1=>2), Flux.flatten, Dense(128=>2), softmax));' 
julia -e   29.41s user 2.05s system 115% cpu 27.155 total

% time julia -e '@time using Flux; @time gradient(m -> Flux.mse(m(rand(Float32,10,10,1,1)), rand(2,1)), Chain(Conv((3,3),1=>2), Flux.flatten, Dense(128=>2), softmax));'
  9.128240 seconds (16.43 M allocations: 1.131 GiB, 3.08% gc time, 47.75% compilation time: 67% of which was recompilation)
 17.482574 seconds (38.85 M allocations: 2.449 GiB, 4.74% gc time, 99.99% compilation time)
julia -e   28.84s user 1.98s system 113% cpu 27.189 total

### With Zygote#master:

% time julia -e '@time using Flux; @time gradient(m -> Flux.mse(m(rand(Float32,10,10,1,1)), rand(2,1)), Chain(Conv((3,3),1=>2), Flux.flatten, Dense(128=>2), softmax));'
  9.281118 seconds (16.90 M allocations: 1.167 GiB, 3.08% gc time, 46.41% compilation time: 66% of which was recompilation)
 12.865140 seconds (17.35 M allocations: 1.074 GiB, 2.66% gc time, 99.99% compilation time)
julia -e   25.27s user 1.89s system 119% cpu 22.726 total

% time julia -e '@time using Flux; @time gradient(m -> Flux.mse(m(rand(Float32,10,10,1,1)), rand(2,1)), Chain(Conv((3,3),1=>2), Flux.flatten, Dense(128=>2), softmax));'
  9.423306 seconds (16.90 M allocations: 1.167 GiB, 3.03% gc time, 46.45% compilation time: 67% of which was recompilation)
 13.019594 seconds (17.35 M allocations: 1.073 GiB, 2.73% gc time, 99.99% compilation time)
julia -e   24.13s user 1.86s system 112% cpu 23.022 total

### Zygote#master  + This PR

% time julia -e '@time using Flux; @time gradient(m -> Flux.mse(m(rand(Float32,10,10,1,1)), rand(2,1)), Chain(Conv((3,3),1=>2), Flux.flatten, Dense(128=>2), softmax));'
 11.086559 seconds (18.63 M allocations: 1.297 GiB, 6.81% gc time, 41.08% compilation time: 66% of which was recompilation)
 12.130469 seconds (7.22 M allocations: 461.883 MiB, 10.48% gc time, 99.99% compilation time)
julia -e   24.53s user 1.50s system 109% cpu 23.814 total

% time julia -e '@time using Flux; @time gradient(m -> Flux.mse(m(rand(Float32,10,10,1,1)), rand(2,1)), Chain(Conv((3,3),1=>2), Flux.flatten, Dense(128=>2), softmax));'
 10.476392 seconds (18.63 M allocations: 1.297 GiB, 3.49% gc time, 41.02% compilation time: 67% of which was recompilation)
 11.062946 seconds (7.22 M allocations: 461.883 MiB, 1.36% gc time, 99.99% compilation time)
julia -e   22.83s user 2.00s system 112% cpu 22.114 total

Cluster xeon with a GPU, Julia 1.8.0

### With Zygote#master

$ time julia-1.8.0/bin/julia -e '@time using Flux; @time gradient(m -> Flux.mse(m(rand(Float32,10,10,1,1)), rand(2,1)), Chain(Conv((3,3),1=>2), Flux.flatten, Dense(128=>2), softmax));'
 21.593748 seconds (22.75 M allocations: 1.485 GiB, 5.96% gc time, 23.77% compilation time: 70% of which was recompilation)
 37.201255 seconds (39.95 M allocations: 2.012 GiB, 4.36% gc time, 99.99% compilation time)

real	0m59.329s
user	1m2.320s
sys	0m2.783s

$ time julia-1.8.0/bin/julia -e '@time using Flux; @time gradient(m -> Flux.mse(m(rand(Float32,10,10,1,1)), rand(2,1)), Chain(Conv((3,3),1=>2), Flux.flatten, Dense(128=>2), softmax));'
 21.987707 seconds (22.75 M allocations: 1.485 GiB, 5.96% gc time, 24.21% compilation time: 69% of which was recompilation)
 37.505275 seconds (39.95 M allocations: 2.011 GiB, 4.39% gc time, 99.99% compilation time)

real	1m0.019s
user	1m2.634s
sys	0m2.701s

### + this PR

$ time julia-1.8.0/bin/julia -e '@time using Flux; @time gradient(m -> Flux.mse(m(rand(Float32,10,10,1,1)), rand(2,1)), Chain(Conv((3,3),1=>2), Flux.flatten, Dense(128=>2), softmax));'
 26.026605 seconds (26.32 M allocations: 1.719 GiB, 5.71% gc time, 20.17% compilation time: 70% of which was recompilation)
 30.415214 seconds (19.08 M allocations: 1002.718 MiB, 3.32% gc time, 99.99% compilation time)

real	0m56.918s
user	0m57.991s
sys	0m2.820s

$ time julia-1.8.0/bin/julia -e '@time using Flux; @time gradient(m -> Flux.mse(m(rand(Float32,10,10,1,1)), rand(2,1)), Chain(Conv((3,3),1=>2), Flux.flatten, Dense(128=>2), softmax));'
 25.925542 seconds (26.32 M allocations: 1.719 GiB, 5.61% gc time, 20.24% compilation time: 70% of which was recompilation)
 30.721970 seconds (19.08 M allocations: 1002.714 MiB, 3.69% gc time, 99.99% compilation time)

real	0m57.128s
user	0m58.368s
sys	0m3.024s

[ToucheSir]

Shall we wait until https://github.com/JuliaLang/julia/pull/47184 lands in nightly to re-evaluate timings? That chain of PRs should reduce the second number significantly. I'm also surprised that timings on the cluster are so slow—maybe something to do with slow file IO?

[mcabbott]

If revived this should probably follow https://github.com/FluxML/Zygote.jl/pull/1415