Clean up Scan's Cython implementation

[brandonwillard]

This PR moves all of the non-native Python, non-NumPy objects out of Scan's Cython implementation so that it has fewer interactions with Python and can—potentially—be turned into a pure C function (i.e. cdef) without too much effort.

Currently, the only function argument that isn't a native or NumPy object is the thunk for the inner-function. When the thunk is a CThunk, we can obtain a Capsule object that I believe Cython cdef functions can use directly; however, that has not been implemented in this PR.

Aside from these simple refactorings, both the Python and Cython implementations have been changed so that they use np.empty to allocate memory. That should provide a small improvement over the current use of np.zeros.

[codecov]

Codecov Report

Merging #774 (bb2ac94) into main (dee152c) will decrease coverage by 0.05%.
The diff coverage is 82.00%.

❗ Current head bb2ac94 differs from pull request most recent head fea7597. Consider uploading reports for the commit fea7597 to get more accurate results

@@            Coverage Diff             @@
##             main     #774      +/-   ##
==========================================
- Coverage   78.40%   78.34%   -0.06%     
==========================================
  Files         152      152              
  Lines       47782    47793      +11     
  Branches    10881    10884       +3     
==========================================
- Hits        37462    37444      -18     
- Misses       7773     7797      +24     
- Partials     2547     2552       +5     

Impacted Files	Coverage Δ
aesara/scan/op.py	81.79% <74.28%> (-0.60%)	⬇️
aesara/compile/function/pfunc.py	83.59% <100.00%> (+0.26%)	⬆️
aesara/scan/basic.py	86.68% <100.00%> (-0.82%)	⬇️
aesara/scan/scan_perform_ext.py	80.32% <100.00%> (ø)
aesara/scan/utils.py	87.44% <100.00%> (+0.01%)	⬆️
aesara/compile/io.py	83.01% <0.00%> (-5.67%)	⬇️
aesara/scan/opt.py	79.48% <0.00%> (-0.84%)	⬇️
aesara/printing.py	48.58% <0.00%> (-0.37%)	⬇️
aesara/tensor/basic_opt.py	85.03% <0.00%> (-0.19%)	⬇️
... and 2 more

[zoj613]

i'm late but I think there is a bit more juice we can squeeze out of scan_perform.pyx as far as minimizing interaction with python objects. The obvious one is using memory views instead off array buffers since they are considered more performant and more flexible (see here). i can investigate a bit and send a PR if this is considered worthy of the effort.

[brandonwillard]

i'm late but I think there is a bit more juice we can squeeze out of scan_perform.pyx as far as minimizing interaction with python objects. The obvious one is using memory views instead off array buffers since they are considered more performant and more flexible (see here). i can investigate a bit and send a PR if this is considered worthy of the effort.

That sounds good.