Fix gene coloring 400 with pandas 3.0 string dtype var index

[colganwi]

Problem

Fixes #2793.

Coloring by a gene returns HTTP 400 for AnnData files whose var index (gene names) uses a pandas string dtype instead of object — the default for text columns under pandas 3.0, and under pandas 2.x with future.infer_string:

GET /api/v0.2/data/var?var:name_0=<gene>  ->  400

Root cause

DataAdaptor._annotation_filter_to_mask chooses between an exact-value filter and a numeric min/max filter by dtype name:

if anno_data.dtype.name in ["boolean", "category", "object"]:
    key_idx = np.in1d(anno_data, values)   # value filter
    ...
else:
    # numeric min/max — a no-op when only "values" is provided

A pandas string dtype ("string", "string[pyarrow]", or the pandas 3.0 NumPy-backed "str") is not in the allow-list, so a gene-name value filter falls into the numeric branch and does nothing. The mask stays all-True, so every var column is selected; data_frame_to_fbs_matrix then exceeds column_request_max and raises ExceedsLimitError → 400. The same applies to obs categorical value filters stored as a string dtype.

Fix

Route all pandas string dtypes through the value-equality branch via pandas.api.types.is_string_dtype:

if anno_data.dtype.name in ["boolean", "category", "object"] or pd.api.types.is_string_dtype(anno_data):

Tests

Adds StringDtypeFilterMaskTest covering both StringDtype ("string") and the pandas 3.0 "str" dtype (via future.infer_string), asserting the value filter selects exactly the matching row.

Reproduction

import anndata, numpy as np, pandas as pd
pd.set_option("future.infer_string", True)  # pandas 3.0 default
adata = anndata.AnnData(
    X=np.random.default_rng(0).random((10, 2000), dtype=np.float32),
    var=pd.DataFrame(index=[f"gene{i}" for i in range(2000)]),
)
adata.write_h5ad("repro.h5ad")
# cellxgene launch repro.h5ad -> color by a gene -> 400 (before this fix)