fix: handle Unicode characters in h5ad metadata without misleading format error

[fresh3nough]

Summary

Fixes #2754

Loading an h5ad file containing non-ASCII Unicode characters (e.g. Greek letter phi in cell labels like Interstitial Mφ perivascular) caused a misleading error: File must be in the .h5ad format. The actual file was valid h5ad, but the broad except ValueError handler in _load_data() discarded the original error context.

Changes

server/data_anndata/anndata_adaptor.py

• Improved ValueError handling in _load_data() to detect Unicode/encoding-related errors and display a meaningful message instead of the generic format error
• Appended the original error message to the generic format error for all other ValueError cases, aiding debugging
• Added _normalize_unicode_strings() static method that normalizes all string metadata in obs/var DataFrames to NFC form during _validate_and_initialize(), preventing encoding issues during serialization

test/unit/data_anndata/test_unicode_metadata.py

• Added regression tests that load h5ad files containing Greek letters, accented characters, and other non-ASCII Unicode in metadata
• Added test verifying NFC normalization is applied to obs metadata strings

Steps to reproduce the original bug

1. Create an h5ad file with a metadata column containing φ (or similar non-ASCII characters)
2. Run cellxgene launch anndata.h5ad
3. Observe misleading error: File must be in the .h5ad format

Steps to verify the fix

python -m unittest test.unit.data_anndata.test_unicode_metadata -v

All 3 new tests pass, and existing data_anndata tests remain unaffected.