fit() crashes with do_approx_sphere=False + pcakeep (PCA-reduced non-symmetric sphere)

[neuromechanist]

Surfaced during #159 review while trying to construct a genuine num_pcs < data_dim test case for AmicaOutput.sources/loadmodout.

Observation

With the Fortran-parity default do_approx_sphere=True, pcakeep < data_dim still yields a full (data_dim, data_dim) (rank-deficient, not shape-reduced) sphere, so num_pcs == data_dim and nothing downstream reduces dimensionality. The only path that actually produces a non-square sphere / num_pcs < data_dim is do_approx_sphere=False + pcakeep, and that currently crashes during fit() itself:

RuntimeError: The size of tensor a (20) must match the size of tensor b (32) at ...

at pyAMICA/torch_impl/core.py:839 (the E-step b = (X - c).T @ W, where the reduced sphere makes the sphered-data channel count disagree with W's dimension).

Why it matters

The genuine PCA-reduction path is unreachable via the public API without hitting this crash, so:

• AmicaOutput.sources/loadmodout/load_results have no test coverage for num_pcs < data_dim (they are correct dimensionally under the symmetric-ZCA default, but the reduced path can't be exercised).
• Any user asking for real PCA reduction with do_approx_sphere=False cannot fit.

Not a regression from #159

Pre-existing; #159 only touches the loader/writer byte order. Filing separately per the parity-regression sequencing rule.

Acceptance

• fit() with do_approx_sphere=False + pcakeep < data_dim runs (sphere/W/c dimensions consistent through the E-step), or rejects the combination up front with a clear error if reduced non-symmetric sphering is intentionally unsupported.
• A round-trip test for sources/loadmodout under genuine num_pcs < data_dim once the fit path works.