Fix figure build failures (M2, M3, M4, MS7)

[aarmey]

Summary

• Move the fancyimpute/scikit-learn check_array compat patch from clustering.py into ddmc/__init__.py. figureM2.py imported fancyimpute before ..clustering, so the patch that swaps force_all_finite for ensure_all_finite (removed in scikit-learn 1.9) was applied too late, and IterativeSVD.fit_transform crashed with TypeError. Package __init__.py always runs before any submodule's own imports, so the fix is now import-order independent.
• Fix get_pssms(clusters=...) misuse in figureM3.py / figureMS7.py: both stripped the cluster axis with a trailing [0], which caused plot_cluster_kinase_distances in common.py to misread the pssm array, pick the wrong "most frequent phosphoacceptor" (falling back to 'A'), filter out all candidate kinases, and crash with an out-of-bounds IndexError. Removed the erroneous [0] and corrected the corresponding indexing in common.py to pssms[i, :, 5] (position-5 column across amino acids).
• Coerce X/y to ndarrays at the top of plot_roc in logistic_regression.py so positional CV-fold indexing (X[train]) works whether the caller passes an ndarray or a pandas Series with a non-integer index — figureM4.py passed a Series and newer pandas raises KeyError instead of silently falling back to positional indexing.

Test plan

• uv run fbuild <name> succeeds for all 11 figures: M2, M3, M4, M5, M6, M7, MS2, MS3, MS4, MS5, MS7
• uv run pytest -q — 21 passed
• uv run ruff check ddmc / uv run ruff format --check clean on touched files
• uv run ty check ddmc passes

🤖 Generated with Claude Code

https://claude.ai/code/session_014sHWCv3yz7DGDVDKra9KUo