Should PyTensor extend its Numba overloads to support all `scipy.sparse.cs[r|c]_matrix` initialization modes?

Description

PyTensor’s Numba overloads for scipy.sparse.cs[r|c]_matrix currently support only a subset of the initialization behaviors provided by the corresponding SciPy constructors.

For example, calling

scipy.sparse.csr_matrix(a_csr_matrix)
scipy.sparse.csc_matrix(a_csc_matrix)

always creates a copy in our Numba overloads, while SciPy allows users to control this behavior via the copy parameter.

I’m opening this issue to:

• Avoid losing track of the discussion started in https://github.com/pymc-devs/pytensor/pull/1860/changes#r2751489368
• Open a broader discussion about whether PyTensor’s Numba overloads for SciPy sparse matrix constructors should support more (or all) of the input combinations and behaviors already supported by SciPy.
  • If we decide to expand support, should this work be done here or in numba-sparse?

Start by reading the Numba overloads for scipy.sparse.csr_matrix and scipy.sparse.csc_matrix, then review the discussion in pull request 1860. Compare the supported initialization modes with the SciPy constructors and clarify whether broader support belongs in PyTensor or numba-sparse. Done means the scope and ownership of the expansion are agreed.