ENH: Allow user configuration of pocketfft's plan caches

[peterbell10]

Reference issue

Closes gh-12297

What does this implement/fix?

This allows the user to interface with scipy.fft's plan caches using 3 utility functions:

• set_plan_cache_size(new_size)
• get_plan_cache_size()
• clear_plan_cache()

Limitations:

• Only controls the number of plans, not amount of memory.
• For simplicity, all plan types have the same size of cache.

Additional information

The C++ implementation introduces two new classes, PlanCache and CacheManager. A plan cache handles LRU caching for a single plan type, similar to what get_plan did before. PlanCache instances are constructed within the CacheManager which keeps an internal reference to every cache. This way I can query and manipulate the caches uniformly by iterating over the CacheManager's internal list of caches.

Note that inside of any given transform, get_plan() interacts directly with the PlanCache and doesn't need to go through the CacheManager. So it exists only for configuration purposes.

TODO:

• Add tests
• Expand documentaion

cc @mreineck, @larsoner

[peterbell10]

I've added a new customisation point, set_plan_cache_max_memsize which limits the maximum memory used for the twiddle factors of a single plan. Anything above that size won't be cached. I've set the default to 2 MB which together with the 16 max plans and the number of internal caches means an absolute upper bound of 576 MB can be taken up by the caches. And to get that to happen you need to use every combination of transform type (fft, rfft, dct and dst) and data type (np.float, np.double, np.longfloat).

[mreineck]

I like the approach, especially the aspect that it comes with a cache size limited by default, so that crashes like the one in #12297 are avoided without users having to learn about the extended interface.

[larsoner]

Agreed this sounds like a good approach