perf: narrow core Polars feature set

[rob-p]

Summary

Replace Polars defaults plus dtype-full with the features used by the core anndata crate.

The retained set covers CSV reading, lazy concatenation, ndarray conversion, categorical values, and the feature-gated 8/16-bit signed and unsigned scalar types supported by AnnData tables. The core crate does not construct Polars temporal, duration, decimal, fixed-array, struct, float16, i128, or u128 columns, and does not need Polars formatting/default bundles.

This keeps existing public data support intact while avoiding unrelated Polars code. The Python crate's separate pyo3-polars dtype requirements are unchanged.

Companion downstream cleanups are COMBINE-lab/grangers#54 and COMBINE-lab/af-anndata#3; the aggregate measurements come from their simpleaf consumer.

Validation

• all 8 anndata unit tests pass
• all 3 anndata doctests pass
• af-anndata builds and passes its unit tests and doctest against this branch
• in the complete simpleaf graph, this and the companion grangers/af-anndata feature changes remove 10 resolved packages; a controlled clean release build was 13.7% faster, used 11.4% less peak compiler memory, and produced an 11.1% smaller unstripped binary

This is intentionally independent of the open Polars 0.55 update and can be rebased onto it if that lands first.