Mainly, joining 0 axes is equivalent to inserting a new dimension. This is the mirror of how splitting a single axis into an empty shape is equivalent to squeezing it.

Also:
* Allow default `None` on unpack

Due to https://github.com/scipy/scipy/issues/24345

Remove cases where type-hints are better than bad type-hints

Inline when only used in one place, or remove if altogether unused

* Do not coerce gradients to TensorVariable

This could cause spurious disconnected errors, because the tensorified variable was not in the graph of the cost

* Type-consistent checks

---------
Co-authored-by: jessegrabowski <jessegrabowski@gmail.com>

This circumvents a bug when DimShuffle of a scalar shows up inside a Blockwise, as the outer indexing yields a float (as opposed to a numpy scalar) which has no `.shape` attribute.

Removes depracated `create_tuple_creator`

* Split blockwise and subtensor tests to deparate CI job

* Skip dimshuffle memory leak test in numba CI

* local_subtensor_of_squeeze bugfix

* Test graph correctness

* Follow template for test

* Remove depreciated AbstractConv Ops and tests

* Remove tensor/conv from test CI

* remove conv.rst

* Add machinery to enable numba caching of function pointers

* Numba cache for Cholesky

* Numba cache for SolveTriangular

* Numba cache for CholeskySolve

* Numba cache for solve helpers

* Numba cache for GECON

* Numba cache for lu_factor

* Numba cache for Solve when assume_a="gen"

* Numba cache for Solve when assume_a="sym"

* Numba cache for Solve when assume_a="pos"

* Numba cache for Solve when assume_a='tri'

* Numba cache for QR

* Clean up obsolete code

* Feedback

* More feedback

* Rename `cache_key_lit` -> `unique_func_name_lit`

* Implement `L_Op` for `join_dims` and `split_dims`

Improve type hints for `join_dims` and `split_dims`

* Feedback