Adaptive-PDF selection for AMICATorchNG

[neuromechanist]

Follow-up from #24 (parity PR #25).

The natural-gradient backend AMICATorchNG matches the Fortran reference with a fixed generalized-Gaussian PDF (component correlation ~0.997). It does not yet implement the adaptive-PDF selection (Laplace / Student-t / GG per source) that the legacy NumPy path and torch_impl/adaptive_pdf.py provide.

This is a feature beyond Fortran parity (the reference binary uses a fixed GG, pdftype 0), so it was intentionally deferred out of the parity work.

Scope:

• Port the adaptive-PDF density/score/curvature into the AMICATorchNG E-step/M-step (_forward, _get_block_updates, _update_parameters).
• Keep the exact-EM + natural-gradient A-update structure from Verify NG init-basin vs Newton bit-parity via Fortran load_* init-matching #24.
• Validate on real sample data; no synthetic data (NO-MOCK).

Reference prototype: pyAMICA/torch_impl/adaptive_pdf.py (used by the parked AMICATorchV2).