Resolve last two open issues: #199 (mypy) and #342 (volumetric)

[zihuaihuai]

Summary

Closes the last two open issues that weren't resolved by #378 / #379 / #380.

Closes #199 — Remove # type: ignore from SLM class

SLM.py and _multiple_comparisons.py had file-level # type: ignore shebangs because mypy crashed on the in-class from ._linear_model import _linear_model pattern (the original blocker, python/mypy#10521). Move those imports to module level and bind the functions as class attributes (_linear_model = _linear_model, etc.). Mypy now runs to completion on both files instead of crashing.

The pre-existing untyped-code errors that the file-level ignore was hiding (~380 of them, mostly around Optional access without narrowing) are now surfaced but left to a follow-up PR — fixing them is genuinely a separate piece of work and re-enabling the commented-out mypy job in CI before that triage would just produce noise.

Closes #342 — Volumetric mixed effects

The reports on this issue all converged on the same two contract bugs in volumetric workflows:

1. contrast dtype: pandas Series of object/categorical dtype was preserved through np.array(contrast). Downstream np.dot then raised TypeError: can't multiply sequence by non-int of type 'float', or (worse) the run silently produced all-NaN t-values. Now coerced to numeric in __init__, with a clear TypeError if conversion fails.
2. mask dtype: integer 0/1 masks (the natural shape of nib.load(...).get_fdata().astype(int)) caused _fdr to do fancy indexing instead of boolean masking, producing the well-known shape (126006,) could not be broadcast to indexing result of shape (1082035,) error. Now coerced to bool in __init__.

Also reworded the surf=None + correction='rft' error so it explicitly points users at correction='fdr' for volumetric data — the previous "RFT corrections require a surface" message read as a dead-end.

The hard part of #342 — implementing cluster-level RFT/TFCE for volumes — is genuinely out of scope; SurfStat had it but BrainStat's port doesn't, and a faithful re-port would be a multi-week project. That's flagged in the new error message and remains tracked. What this PR does fix: every SLM(...) invocation pattern that the issue reporters posted now runs end-to-end.

Test plan

Four new tests in brainstat/tests/test_SLM.py:

• test_volumetric_contrast_coercion — pandas Series → float; object dtype rejected with helpful message.
• test_volumetric_int_mask_is_coerced_to_bool — int 0/1 mask becomes a boolean mask.
• test_volumetric_rft_without_surface_explains_alternatives — error message mentions FDR.
• test_volumetric_fdr_with_int_mask_runs_end_to_end — the full reporter scenario (int mask + pandas contrast + FDR) returns finite t-values inside the mask and a Q-array of the right shape.

Locally: 73/73 SLM + t_test + FDR + f_test cases pass.

Out of scope (kept open as separate trackers)

• The 383 pre-existing typing errors that the file-level # type: ignore was hiding — straightforward but noisy cleanup.
• Volumetric cluster-level RFT/TFCE (the ambitious half of Volumetric mixed effects #342).
• Re-enabling the commented-out mypy job in python_unittests.yml once the typing errors are triaged.