Implement vectorized adstock transformations

[ricardoV94]

Follow up to #114

Unlike #114 it doesn't try to respect broadcastable flags as those are in a limbo in PyTensor at the moment (see pymc-devs/pytensor#149) and are not respected anywhere else.

Closes #22
Closes #196

[cluhmann]

Ok. Can verify this works as (I) expected in #196. Will review now!

[cluhmann]

This looks good to me, though the details of batched_convolution are pretty opaque. @lucianopaz should definitely review that as it was lifted from his PR.

[lucianopaz]

I suggest adding a bunch of comments to make the convolution code a bit easier to read

[ricardoV94]

Thanks @lucianopaz!

[codecov]

Codecov Report

Merging #221 (b341199) into main (edf0530) will decrease coverage by 0.57%.
The diff coverage is 93.54%.

❗ Current head b341199 differs from pull request most recent head 93f2ac8. Consider uploading reports for the commit 93f2ac8 to get more accurate results

@@            Coverage Diff             @@
##             main     #221      +/-   ##
==========================================
- Coverage   94.83%   94.27%   -0.57%     
==========================================
  Files          18       18              
  Lines         968      978      +10     
==========================================
+ Hits          918      922       +4     
- Misses         50       56       +6     

Impacted Files	Coverage Δ
pymc_marketing/mmm/transformers.py	94.44% <93.33%> (-5.56%)	⬇️
pymc_marketing/mmm/delayed_saturated_mmm.py	91.66% <100.00%> (-8.34%)	⬇️

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

[aseyboldt]

@ricardoV94 Didn't this PR change the runtime of the geometric adstock to something slower (from $O(n)$ to $O(n n_\text{lag})$)?
I didn't perform any testing, and maybe on a GPU it doesn't make as much difference, but at least for longer lag times this should be worse?

[ricardoV94]

@ricardoV94 Didn't this PR change the runtime of the geometric adstock to something slower (from O(n) to $O(n n_\text{lag})$)? I didn't perform any testing, and maybe on a GPU it doesn't make as much difference, but at least for longer lag times this should be worse?

Possibly, the concern was vectorization not speed. I didn't write the code, it was actually @lucianopaz, this was a small change from #114

[aseyboldt]

Makes sense. I guess we can just try to remember in case we run into performance issues :-)

[williambdean]

This came up recently #670

[twiecki]

Could we keep both implementations?