Also cleans up implementation and documentation

* Split and inverse

* PyTorch inline constants in dispatch to avoid graph breaks

Deprecate obscure ndim kwarg

A shortcut in the numpy implementation of einsum_path when there's nothing to optimize, creates a default path that can combine more than 2 operands. Our implementation only works with 2 or 1 operand operations at each step.

https://github.com/numpy/numpy/blob/cc5851e654bfd82a23f2758be4bd224be84fc1c3/numpy/_core/einsumfunc.py#L945-L951

Nothing clever was going on, unknown dims were simply faked as having length 8 for JAX polymorphism export

Internal API changed in numba 0.61

Existing benchmarks don't show any difference in performance

It replaces the old warning that does not actually apply by a more informative and actionable one.

This warning was for Ops that might use the alternative blas_headers, which rely on the Numpy C-API.

However, regular PyTensor user has not used this for a while. The only Op that would use C-code with this alternative headers is the GEMM Op which is not included in current rewrites. Instead Dot22 or Dot22Scalar are introduced, which refuse to generate C-code altogether if the blas flags are missing.

* Add 'overwrite_existing' flag to allow graph rewrites
and include appropriate testing

* Encapsulate test rewriters and use user-facing API

---------
Co-authored-by: Ricardo Vieira <ricardo.vieira1994@gmail.com>

Also allow arbitrary capitalization of the modes.

Also make linker and optimizer non-mutable config as the mode is cached after using them for the first time.

Accidentally introduced in c655b028

Also move tests to the rewriting test file

* Add trusted publishing

* Don't persist credentials in pypi.yml

Ref: <https://woodruffw.github.io/zizmor/audits/#artipacked>

* Don't attempt to publish from forks

* Include more comments

* Don't persist credentials in pypi.yml

Ref: <https://woodruffw.github.io/zizmor/audits/#artipacked>

* Don't persist credentials

This is an insecure default on GitHub that increases the chances of credential leakage.
<https://unit42.paloaltonetworks.com/github-repo-artifacts-leak-tokens/>

* Add zizmor exception for RTD pull_request_target trigger

* Create zizmor workflow

1. Use actual Solve Op to infer output dtype as CholSolve outputs a different dtype than basic Solve in Scipy==1.15

2. Tweaked test related to https://github.com/pymc-devs/pytensor/issues/1152

3. Tweak tolerage

Fixes bug in `local_add_neg_to_sub` reported in https://github.com/pymc-devs/pytensor/issues/584