This reverts commit 69319ef8795712427d41902e4fcf24260af716df.

Bumps [pypa/cibuildwheel](https://github.com/pypa/cibuildwheel) from 2.14.1 to 2.16.5.
- [Release notes](https://github.com/pypa/cibuildwheel/releases)
- [Changelog](https://github.com/pypa/cibuildwheel/blob/main/docs/changelog.md)
- [Commits](https://github.com/pypa/cibuildwheel/compare/v2.14.1...v2.16.5)

---
updated-dependencies:
- dependency-name: pypa/cibuildwheel
  dependency-type: direct:production
  update-type: version-update:semver-minor
...
Signed-off-by: dependabot[bot] <support@github.com>

In some environments, `sys.stdout.encoding` may return `None`. 

This is encountered when knitting an rmarkdown document using the reticulate engine. In that context, `sys.stdout` is replaced with a `StringIO()` object to capture all output. `StringIO().encoding` returns `None`

closes https://github.com/rstudio/reticulate/issues/1508

updates:
- [github.com/astral-sh/ruff-pre-commit: v0.1.13 → v0.1.14](https://github.com/astral-sh/ruff-pre-commit/compare/v0.1.13...v0.1.14)

Also make test slightly more restrictive

* Implement numba overload for POTRF, LAPACK cholesky routine

* Delete old numba_funcify_Cholesky

* Refactor tests to include supported keywords and datatypes

* Validate inputs and outputs of numba cholesky function

* Raise on complex inputs

* Change `cholesky` default for `check_finite` to `False`

* Remove redundant dtype checks from numba linalg dispatchers

* Add docstring to `numba_funcify_Cholesky` explaining why the overload is necessary.

Co-authored-by: Ben Mares <services-git-throwaway1@tensorial.com>

Makes downstream packaging simpler and more robust, avoiding the need
for downstream patches when versioneer is updated.

Building with versioneer v0.29.0 works fine on Arch Linux, see:

  https://aur.archlinux.org/packages/python-pytensor

* Replace mention of Aesara with Theano as Aesara is discontinued and Theano is the better-known library.

* Also remove fork of Theano.

* Mention PyMC.

When replacing the Shared RNG variables, the input order of the FunctionGraph was not explicitly aligned with the input storage of the function being compiled.

The provided output could be silently ignored and replaced by the new output of the vectorized node.

The changes also avoid vectorizing multiple-output nodes when none of the unreplaced outputs are needed.

* Copy `block_diag` and support functions from `pymc.math`

* Evaluate output in sphinx code example
Co-authored-by: Ricardo Vieira <28983449+ricardoV94@users.noreply.github.com>

* Test type equivalence with `isinstance` instead of `==`
Co-authored-by: Ricardo Vieira <28983449+ricardoV94@users.noreply.github.com>

* Typo in test function

* Split `block_diag` into sparse and dense version

Closely follow scipy function signature for `block_diag`

* Use `as_sparse_or_tensor_variable` in `SparseBlockDiagonalMatrix` to allow sparse matrix inputs to `pytensor.sparse.block_diag`

* Test sparse and dense inputs to `pytensor.sparse.block_diag`

* Add numba overload for `pytensor.tensor.slinalg.block_diag`

* add jax overload for `pytensor.tensor.slinalg.block_diag`

* Move stand-alone `block_diag_grad` function into `grad` method

* Add `format` prop to `SparseBlockDiagonalMatrix`

* Use `compare_numba_and_py` in `numba\test_slinalg.py::test_block_diag`

* Add support for Blockwise to `slinalg.block_diag`

* Add gradient test

Remove `Matrix` from `BlockDiagonal` and `SparseBlockDiagonal` `Op` names

Correct errors in docstrings

Move input validation to a shared class method

* Remove `gufunc_signature` from `__props__`
Co-authored-by: Ricardo Vieira <28983449+ricardoV94@users.noreply.github.com>

* Implement correct `__props__` for subclasses of `BaseBlockMatrix`

---------
Co-authored-by: Ricardo Vieira <28983449+ricardoV94@users.noreply.github.com>

* Use `solve_triangular` instead of in `psd_solve_with_chol`

* Add unittest for `psd_solve_with_chol`

* Specify `mode=FAST_RUN` in test

* Relax `test_psd_solve_with_chol` `atol` and `rtol` for half-precision tests

Also refactors shared logic for other batch axed Ops

All types of axis arguments are supported by max_and_argmax

This reverts commit 0fc2cd8e.

This rewrites replaces a batched matmul by a core matmul by raveling the batched dimensions of the batched input, doing a 2D matmul and then unravelling the batched dimensions of the output.

This sidesteps the Blockwise Dot operation and allows specialization into BLAS routines.

The idea was taken from these two discussions:
https://github.com/numpy/numpy/issues/7569
https://github.com/numpy/numpy/issues/8957