Compute average_batch_r2 per cell type

[rcannood]

Describe your changes

average_batch_r2 reports a per-cell-type score (average_batch_r2_ct), but the R² it computes is the donor-level one, pooled over all cell types.

In src/metrics/average_batch_r2/script.py the loop builds s1_view_ct / s2_view_ct, uses them for the 20-cell guard, and then hands batch_r2() the donor-level s1_view / s2_view:

for ct in ct_list:
    s1_view_ct = s1_view[s1_view.obs["cell_type"] == ct]
    s2_view_ct = s2_view[s2_view.obs["cell_type"] == ct]
    if s1_view_ct.shape[0] < 20 or s2_view_ct.shape[0] < 20:
        continue

    r2_list, marker_list = batch_r2(s1_view, s2_view)   # <-- not the _ct views

Two consequences:

• The reported value is the global R², not the cell-type-aware one -- so it can't show the "good on global but not on cell type" case that the metric description is built around, and the ∑_{j=1}^{C} cell type sum in the description isn't what's computed.
• Because the same donor-level value is appended once per surviving cell type, donors with more cell types above the 20-cell threshold get more weight in the final np.mean().

This PR passes the cell type subsets, and corrects the "less than 10 cells" comment next to the < 20 guard.

Note that this changes the value of average_batch_r2_ct for every method, so results from earlier runs are not comparable.

Found while reviewing the task ahead of the next full benchmark run -- see also #TODO_ISSUE_A and the sibling PRs.

Checklist before requesting a review

• I have performed a self-review of my code

• Check the correct box. Does this PR contain:

  • Breaking changes
  • New functionality
  • Major changes
  • Minor changes
  • Bug fixes

• Proposed changes are described in the CHANGELOG.md

• CI Tests succeed and look good!

[LuLeom]

LGTM