Obs filter

[jjschirle]

Adds optional obs-based cell subsetting to load_and_normalize() via an obs_filter callable.

• bumped python version pinned to 3.12 as required in pyproject.toml.
• Adds obs_filter, e.g. obs_filter=lambda obs: obs["condition"] == "control".
• Applies obs subsetting before cell filtering, gene filtering, depth normalization, and log transformation.
• Computes gene totals and the gene_threshold cutoff using only the selected cohort.
• Preserves the existing load_and_normalize() implementation when obs_filter=None.
• Avoids materializing decoded indices and values for excluded IVCSR rows. The packed byte stream must still be read and scanned because byte offsets for major slices are not currently stored.
• Adds validation for invalid masks, empty selections, non-callable filters, and files without obs.
• Adds tests comparing filtered results against preprocessing an equivalent AnnData subset, including interleaved conditions and metadata slicing.

The callable is easily extended:

obs_filter=lambda obs: (
    obs["condition"].isin(["control", "vehicle"])
    & obs["timepoint"].isin(["T1", "T3"])
    & (obs["batch"] != "bad_batch")
)

Passes all rapid_load pytests, ruff, and ty checks.