Bug was caused by reusing the adjusted strides in the logic to decide whether the call to GEMV should be transposed or not.

Particularly the +1 in the strides variable was causing the error branch (no double-strides) to be reached wrongly. The +1 was supposedly there for the case of matrix with length 0, but that triggers a branch where the adjusted strides are never used.

This bug was introduced in afe934b2

Bumps [pypa/cibuildwheel](https://github.com/pypa/cibuildwheel) from 2.23.0 to 2.23.3.
- [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/v2.23.0...v2.23.3)

---
updated-dependencies:
- dependency-name: pypa/cibuildwheel
  dependency-version: 2.23.3
  dependency-type: direct:production
  update-type: version-update:semver-patch
...
Signed-off-by: dependabot[bot] <support@github.com>

Set view flag of ExtractDiagonal to True and respect by default

Split off Subtensor of Unbroadcast into its own rewrite

It seems to add a non-trivial 100ns

Signed-off-by: Emmanuel Ferdman <emmanuelferdman@gmail.com>

Also add checks for runtime broadcast

These show up in the gradient of Convolve1D

Such that its name can be more informative.

They could get tested independently.
This was mostly done to understand their purpose.