Replace fslmaths calls with nibabel + numpy/scipy

[nx10]

All 8 fslmaths calls and 1 3dcalc call in the codebase are simple voxelwise or morphological operations that can be done with nibabel + numpy/scipy, removing container calls entirely.

functional/masking.py (6 calls)

In bold_masking(), the entire mask refinement chain is fslmaths:

1. thr + bin — threshold warped probseg at 0.85, binarize
2. kernel_sphere + dil_f — dilate binary mask by 3mm sphere
3. kernel_sphere + dil_f — dilate BET mask by 6mm sphere
4. mas — apply dilated BET mask to bias-corrected BOLD ref
5. mul — intersect BET mask with AFNI automask
6. mas — apply final mask to unifized image

All of these are one-liners with numpy (>, *, astype) or scipy (binary_dilation with ball). We already use binary_erosion in distortion.py for the same kind of thing.

anatomical/segmentation.py (2 calls)

1. thr + bin in fsl_tissue_masks() — threshold FAST PVE at 0.95, binarize
2. thr + uthr + bin in fsl_wm_bbr_mask() — band-threshold FAST pveseg [2.5, 3.5], binarize

Same story, pure numpy.

functional/initialization.py (1 call)

1. 3dcalc in truncate_trs() — selects a[start_tr..$] with expression a. Just img.slicer[..., start_tr:] in nibabel.

Why bother

• Removes 9 container round-trips (meaningful in Docker mode)
• Lets us write intermediates as uncompressed .nii without fighting FSLOUTPUTTYPE (see Consider disabling gzip compression for certain nifti intermediates #252 / Disable gzip compression for float NIfTI intermediates #285)
• Simpler code, fewer dependencies on FSL/AFNI for pure math