Numba: SIMD-friendly maximum/minimum

[ricardoV94]

Closes #2367

This PR goes from 5x to 1.1-1.3x slower than numpy on my machine. I don't think we can go below that because numba/llvm won't SIMD these operations automatically. Numpy has a SIMD dispatcher instead.

In my experience this is not super critical as max/min are usually fused in loops that are order of magnitude worse, usually exp in a stable softmax or logsumexp

[jessegrabowski]

For the record it came up in maxpool layers here: pymc-devs/pytensor_ml#100

[ricardoV94]

For the record it came up in maxpool layers here: pymc-devs/pytensor_ml#100

Does it fix it on your end?

[jessegrabowski]

I switched to a custom kernel for those layers so it was moot, I just raised the issue because i found it was slow.

I'll re-try the naive path tonight and let you know

[ricardoV94]

I switched to a custom kernel for those layers so it was moot, I just raised the issue because i found it was slow.

I'll re-try the naive path tonight and let you know

If the layer is literally a max it's probably one of those you'd want to fuse with whatever produces it