Fix prot.pp.clr for old-style sparse matrices

[Zethson]

Problem

muon.prot.pp.clr crashes on the seurat flavor when adata.X is an old-style SciPy sparse matrix (csc_matrix / csr_matrix):

mu.prot.pp.clr(adata, axis=1)
# AttributeError: 'matrix' object has no attribute 'toarray'

For an spmatrix, np.log1p(x).mean(axis=axis) returns a numpy.matrix, which has no .toarray() method, so the existing line fails. This affects common datasets whose ADT .X is still an spmatrix (e.g. papalexi_2021).

Fix

Compute the per-feature log-mean once and coerce it to a plain array depending on the container type:

• old-style csc_matrix / csr_matrix: use .A and x.getnnz(axis=axis)
• sparse arrays: use np.diff(x.indptr) to avoid the getnnz behavior change (ENH: sparse.csr_array: keep axis functionality in getnnz() scipy/scipy#19405)

The axis type hint is also tightened to Literal[0, 1].

Verification

Installed this branch and ran the pertpy perturbation_efficacy tutorial (which calls mu.prot.pp.clr(mdata["adt"], axis=1) on papalexi_2021) end-to-end: all 32 code cells execute with no errors. CLR output on both axis=0 and axis=1 is finite, non-negative, and preserves sparsity as expected for the seurat flavor.