fix(nifti): enable lazy loading for Nifti1ImageWrapper

[The-Obstacle-Is-The-Way]

Summary

• Single-line fix: Change dataobj=nifti_image.get_fdata() → dataobj=nifti_image.dataobj
• Preserves nibabel's ArrayProxy for true lazy loading instead of eagerly loading entire NIfTI files into memory
• Improves error handling: corrupted files now fail at access time with clear context instead of silently during decode

Problem

The Nifti1ImageWrapper.__init__ was calling get_fdata() which immediately loads the entire image into memory. For large 4D fMRI files (often 1-2GB), this causes:

1. Memory issues - Full data loaded during decode, not on demand
2. Poor error handling - Corrupted files crash at access time with unclear error messages (e.g., EOFError with no file path)
3. No graceful recovery - Entire dataset iteration fails on one bad file

Solution

Use nifti_image.dataobj which preserves the underlying ArrayProxy, deferring actual I/O to get_fdata() calls.

Test Plan

• Added test_nifti_lazy_loading to verify ArrayProxy is preserved
• All 22 existing NIfTI tests pass
• End-to-end tested with real OpenNeuro data (ds000102)
• CodeRabbit approved: "Switch to dataobj correctly restores nibabel's lazy loading semantics... This looks solid"

Related

• Discovered while testing BIDS loader PR feat(bids): Add BIDS dataset loader for neuroimaging data #7886
• Complements the NIfTI + NiiVue viewer work from Replace papaya with niivue #7878 and Nifti visualization support #7874