Use set_valid_worlds! instead of manual optimisation

[mhauru]

Replicating the changes made to Mooncake by in chalk-lab/Mooncake.jl#832.

The benchmarks that pass, run on this branch:

benchmarking rosenbrock...
  Run Original Function:  185.542 μs (24 allocations: 6.25 MiB)
  Run TapedTask: #produce=1;   6.784 ms (299145 allocations: 10.83 MiB)
benchmarking ackley...
  Run Original Function:  1.154 ms (0 allocations: 0 bytes)
  Run TapedTask: #produce=100000;   60.412 ms (899584 allocations: 21.36 MiB)
benchmarking matrix_test...
  Run Original Function:  97.417 μs (18 allocations: 576.47 KiB)
  Run TapedTask: #produce=1;   468.625 μs (530 allocations: 594.08 KiB)
benchmarking neural_net...
  Run Original Function:  445.707 ns (8 allocations: 576 bytes)
  Run TapedTask: #produce=1;   12.625 μs (168 allocations: 6.27 KiB)

on main:

benchmarking rosenbrock...
  Run Original Function:  169.583 μs (24 allocations: 6.25 MiB)
  Run TapedTask: #produce=1;   2.062 ms (679 allocations: 6.27 MiB)
benchmarking ackley...
  Run Original Function:  1.154 ms (0 allocations: 0 bytes)
  Run TapedTask: #produce=100000;   56.672 ms (699582 allocations: 18.31 MiB)
benchmarking matrix_test...
  Run Original Function:  98.250 μs (18 allocations: 576.47 KiB)
  Run TapedTask: #produce=1;   294.750 μs (529 allocations: 594.06 KiB)
benchmarking neural_net...
  Run Original Function:  430.065 ns (8 allocations: 576 bytes)
  Run TapedTask: #produce=1;   12.041 μs (168 allocations: 6.27 KiB)

and on main but skipping the manual optimisation and just calling misty_closure directly:

benchmarking rosenbrock...
  Run Original Function:  173.083 μs (24 allocations: 6.25 MiB)
  Run TapedTask: #produce=1;   6.428 ms (299145 allocations: 10.83 MiB)
benchmarking ackley...
  Run Original Function:  1.153 ms (0 allocations: 0 bytes)
  Run TapedTask: #produce=100000;   61.659 ms (899584 allocations: 21.36 MiB)
benchmarking matrix_test...
  Run Original Function:  98.542 μs (18 allocations: 576.47 KiB)
  Run TapedTask: #produce=1;   389.250 μs (530 allocations: 594.08 KiB)
benchmarking neural_net...
  Run Original Function:  435.187 ns (8 allocations: 576 bytes)
  Run TapedTask: #produce=1;   11.791 μs (168 allocations: 6.27 KiB)
done

Two of these are close, but the Rosenbrock one is 230% slower and matrix_test about 60% slower on this branch, and very similar to if I just disable the manual optimisation.

[github-actions]

Libtask.jl documentation for PR #205 is available at:
https://TuringLang.github.io/Libtask.jl/previews/PR205/

[mhauru]

After some optimisations provided by @Technici4n, the benchmarks on this branch are now on par with main:

benchmarking rosenbrock...
  Run Original Function:  193.709 μs (24 allocations: 6.25 MiB)
  Run TapedTask: #produce=1;   1.871 ms (678 allocations: 6.27 MiB)
benchmarking ackley...
  Run Original Function:  1.159 ms (0 allocations: 0 bytes)
  Run TapedTask: #produce=100000;   53.375 ms (500092 allocations: 15.26 MiB)
benchmarking matrix_test...
  Run Original Function:  99.625 μs (18 allocations: 576.47 KiB)
  Run TapedTask: #produce=1;   295.333 μs (528 allocations: 594.05 KiB)
benchmarking neural_net...
  Run Original Function:  439.394 ns (8 allocations: 576 bytes)
  Run TapedTask: #produce=1;   11.916 μs (168 allocations: 6.27 KiB)

That makes this ready for review.

[sunxd3]

looks clean

[sunxd3]

great work, both!

[mhauru]

This really all @Technici4n, I just copied over some of his code from Mooncake/Slack and added comments. Thank you!

[penelopeysm]

I think something here broke Turing: https://github.com/TuringLang/Turing.jl/actions/runs/19153277686/job/54748336850?pr=2710