Drop Aesara support and fix PyTensor compatibility

[michaelosthege]

This simplifies testing & upcoming fixes related to changes in shape handling in PyTensor.

[michaelosthege]

@ricardoV94 @aseyboldt the tests here are failing with this:

E                           pytensor.graph.utils.MissingInputError: Input 0 (beta) of the graph (indices start from 0), used to compute Shape(beta), was not provided and not given a value. Use the PyTensor flag exception_verbosity='high', for more information on this error.

But the shape of beta should be trivial, and already known because it's scalar!

beta = pt.dscalar("beta")

Am I missing something?

[ricardoV94]

We keep the shape dependencies since pymc-devs/pytensor#280

This allows us to potentially clone_replace(..., strict=False), so you could change a scalar by a vector (clone_replace is a bit crude, but the idea is that).

As a solution you can use pymc.pytensorf.constant_fold to get rid of useless shape Ops if you don't want to those inputs to show up in the compiled graphs.

[michaelosthege]

Oh, so already this no longer works:

pt.dscalar().shape.eval()

But now I'm maximally confused

[ricardoV94]

There's a distinction between x.shape (pytensor Op) and x.type.shape (python tuple). The "static shape" refers to the latter and that's what is being shown in the __repr__

[michaelosthege]

The inspiration for the NumPy minimum version bumps came from conda/conda-build#4498 but it didn't help.

The successful 3.8 jobs were using an old PyTensor version.. Let's see what Python 3.11 brings

[maresb]

Ah, cool! I did that last night and didn't realize it succeeded! 🚀

Unless sunode has some explicit numpy version requirements then it's best to pin implicitly.