lightweight random greedy optimization

[jcmgray]

This adds support for lightweight random greedy optimization (no hyper parameter tuning or ContractionTree construction), with the following:

• optimize_random_greedy_track_flops, with a cotengrust version to follow.
• RandomGreedyOptimizer, a stateful optimizer which wraps the python or rust version above, with parallelization etc.

TODO:

• maybe a ReusableRandomGreedyOptimizer

cc @CSNWEB

[pep8speaks]

Hello @jcmgray! Thanks for updating this PR. We checked the lines you've touched for PEP 8 issues, and found:

• In the file cotengra/pathfinders/path_basic.py:

Line 795:80: E501 line too long (82 > 79 characters)

Comment last updated at 2024-04-19 20:20:05 UTC

[CSNWEB]

Awesome that was fast, thanks! I will have a look at it later this week.

[jcmgray]

Rust counterpart is here: jcmgray/cotengrust@f3a3e50.

With the overall performance increase (and memory decrease), pretty big contractions are now amenable:

import cotengra as ctg

# 10,000 nodes
inputs, output, shapes, size_dict = ctg.utils.lattice_equation([100, 100])
opt = ctg.RandomGreedyOptimizer(max_repeats=128)

%%timeit
ssa_path = opt.ssa_path(inputs, output, size_dict)
# 1.75 s ± 48.4 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

[jcmgray]

Merging for the moment, but open to any additional suggestions before releasing it.

[CSNWEB]

Thanks again, looked through the code and it looks very good to me. Seems to contain everything I mentioned and could need.