Module-level Op instances break LSP signatures, Sphinx docs, and source links for expm, det, inv, lstsq

[williambdean]

Problem

Several public API symbols in pytensor.tensor.linalg are bare Op instances (or Blockwise(Op()) instances), not functions with real signatures. This causes three issues:

1. LSPs (pyright, mypy, jedi) can only show the generic Op.__call__ signature (*inputs, name=None, return_list=False, **kwargs) — no parameter names, types, or docstrings.
2. Sphinx autodoc renders the same generic signature and pulls the docstring from Blockwise instead of the actual operation.
3. "Source" links in the docs point to pytensor/tensor/blockwise.py rather than the operation's actual home (e.g. products.py, summary.py, inverse.py).

Affected symbols:

Symbol	File	Current definition	Docs source link points to
expm	products.py:81	expm = Blockwise(Expm())	blockwise.py
det	summary.py:81	det = Blockwise(Det())	blockwise.py
inv, matrix_inverse	inverse.py:166	inv = matrix_inverse = Blockwise(MatrixInverse())	blockwise.py
lstsq	solvers/lstsq.py:36	lstsq = Lstsq()	solvers/lstsq.py

Documentation examples:

• https://pytensor.readthedocs.io/en/latest/library/tensor/linalg.html#pytensor.tensor.linalg.expm — shows expm(*inputs, name=None, return_list=False, **kwargs) with the Blockwise docstring, source link goes to blockwise.py
• https://pytensor.readthedocs.io/en/latest/library/tensor/linalg.html#pytensor.tensor.linalg.det — same issue
• https://pytensor.readthedocs.io/en/latest/library/tensor/linalg.html#pytensor.tensor.linalg.inv — same issue

Compare with properly wrapped functions like solve, cholesky, svd, norm, tensorsolve which all show their actual parameter signatures, domain-specific docstrings, and correct source links.

Possible approaches

A) Function wrapper — The codebase already uses this pattern in tensor/math.py:

_matmul = Blockwise(_dot, name="Matmul")      # private Op instance

def matmul(x1, x2, dtype=None):                 # public function with real signature
    """Compute the matrix product of two tensor variables.
    ...
    """
    return _matmul(x1, x2)

For each affected symbol, this would mean keeping the Op instance private and exposing a wrapper function with proper annotations and docstring.

B) .pyi stub files — Add products.pyi, summary.pyi, etc. with typed signatures. Helps LSPs but doesn't fix Sphinx autodoc, and adds maintenance burden keeping stubs in sync.

C) typing.cast with a Callable annotation — Helps LSPs but doesn't fix Sphinx.

D) Instance __doc__ — e.g. a docstring after the assignment. Fixes the rendered description in Sphinx but not the signature, source link, or LSP behavior.

[ricardoV94]

I'm okay with function wrapper, as long as we don't duplicate docstrings across multiple places, place them only on the user facing functions