Add way to set backend fn random generators

[lucianopaz]

This PR depends on #7540. Do not merge before that one has been merged.

Description

This PR adds the set_function_rngs function that takes a compiled pytensor function, looks for any random generators in it, and makes a copy that sets all generators to spawned versions of a supplied generator. The implementation is taken from @ricardoV94's function here (and that's why I listed him as coauthor of this commit). This function is then used in init_traces and base trace initialization to make the backend's fn have a properly seeded random generator. Thanks to this change, we can reproducible sampling results even when the function has Deterministic that depend on raw pytensor random variables.

Related Issue

• Closes trace fn does not respect random seed for stochastic "Deterministics" #7588
• Related to #

Checklist

• Checked that the pre-commit linting/style checks pass
• Included tests that prove the fix is effective or that the new feature works
• Added necessary documentation (docstrings and/or example notebooks)
• If you are a pro: each commit corresponds to a relevant logical change

Type of change

• New feature / enhancement
• Bug fix
• Documentation
• Maintenance
• Other (please specify):

[ricardoV94]

Need to raise an error if the linker is JAXLinker as function.get_shared() doesn't retrieve the shared RNGs that are actually used by JAX functions (they are incompatible with numpy Generators): https://github.com/pymc-devs/pymc-extras/blob/c1809e8149fc89ac6eadf4bf73050ea6fe82955c/pymc_experimental/sampling/optimizations/conjugate_sampler.py#L79-L82

[ricardoV94]

Perhaps a more explicit name like copy_function_with_new_rngs? That gives the mental model that the old function is not affected?