Add fix_rbc_release_tr.py script for TR bandpass bug

Lets users of an RBC data release re-derive a correctly-bandpassed
cleaned BOLD and all downstream metrics, by patching the template-space
`desc-head_bold.nii.gz` header (TR=0.0 on disk) with the TR read from
the native-space `desc-preproc_bold.nii.gz` before regression.

Outputs land under a parallel `--output-dir` tree using the published
release's filename conventions:

  * cleaned BOLD + bandpass-filtered regressors
  * ALFF / fALFF / ReHo at `desc-sm6`, `desc-smZstd`, `desc-zstd`
  * Atlas-based mean timeseries (`.1D`, `(n_timepoints, n_rois)`)
  * Pearson + nilearn partial-correlation matrices, per atlas

Metrics compose `rbc.workflows.metrics.single_session_metrics` and
`rbc.core.metrics.standardization.compute_zscore`. Atlas resolution
goes through `rbc_resources.ATLAS_REGISTRY` so the release-name -> file
mapping survives filename changes inside `rbc_resources`.

The script ships with PEP 723 inline metadata so it can be invoked
standalone via `uv run --script` without a clone. `--skip-metrics`
falls back to BOLD-only behavior. `--verify` scans every discoverable
run (header-only reads) and reports buggy / clean / inconclusive
counts. `RBC_WORK_DIR` is scoped to the script's tempdir for the
duration of the run so all niwrap exec folders get cleaned up.

Also promotes `_restore_tr` to a public `restore_tr` for reuse, and
adds `nilearn` to the dev dependency group (used by the partial
correlation step; declared in PEP 723 deps for standalone use).

Author: nx10