Return NaN instead of crashing when every EMD stratum is too small

[rcannood]

Describe your changes

Both EMD paths skip a stratum when either side has fewer than 50 cells, and both then call pd.concat() on the collected results. If every stratum is skipped, that list is empty and pd.concat() raises ValueError: No objects to concatenate, so the whole metric component dies instead of reporting that it couldn't be computed.

The vertical path already handles the neighbouring case gracefully -- if there aren't at least 2 samples per group it returns np.nan -- so this PR extends the same treatment to the empty-results case in both paths:

• get_vert_emd_for_integrated_adata() returns np.nan when no sample combination and cell type had enough cells.
• calculate_horizontal_emd() returns np.nan for both its keys when no donor and cell type had enough cells.

Also two small things in the same block:

• The "some cell types are in split 1 but not in split 2" warning was gated on len(cell_type_not_in_both) > 1, so it stayed quiet in the most common case of exactly one missing cell type. Changed to > 0.
• That warning's message ended in ]n rather than a newline, and joined the cell type names without a separator.

Note that with errorStrategy set to ignore after retries, a crash here currently looks the same as a missing score in the output -- which is part of why it is worth turning into an explicit NaN.

Found while reviewing the task ahead of the next full benchmark run -- see also 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!

[ghar1821]

Not fixing the emd crashing on pd.concat() when every donor/sample combination and cell type falls below the 50 cell threshold, because we would have detected this prior to running any methods or metrics, and we would not have included the dataset if this is the case (PR #131).