feat: bump polars 0.53 -> 0.55.2

[rob-p]

Closes #39.

Stacked on #40. The first commit here is that PR's one-line build fix, without which anndata-hdf5 does not compile at all and CI cannot go green. Merge #40 first and I will rebase, or take both together — the polars work is the second commit.

Why

polars is pre-1.0, so each minor release is a distinct, incompatible set of types. anndata takes and returns polars::prelude::DataFrame across its public API but does not re-export polars, so every downstream crate must declare a polars dependency matching this one exactly. That means this pin decides the polars version for the whole downstream ecosystem, and it is currently two minor versions behind.

What changed

crate	from	to	why
polars	0.53	0.55.2	the actual goal
pyo3-polars	0.26	0.28	first release supporting polars 0.55
pyo3	0.27	0.29	required by pyo3-polars 0.28
numpy	0.27	0.29	tracks pyo3 in lockstep
pyo3-log	0.13	0.13.4	earlier releases cap at pyo3 < 0.29

Two things worth knowing:

• The requirement has to be 0.55.2, not 0.55. polars 0.55.0 declares arrow = "^0.55.0" and polars-parquet = "^0.55.0", but polars-parquet 0.55.2 requires polars-arrow ^0.55.2, so a resolution that picks 0.55.0 fails with failed to select a version for polars-arrow.
• pyo3 0.29 required no source changes. The bump is purely a manifest change on that side.

The only code change is that polars 0.55 removed impl IntoIterator for &ChunkedArray<T>, so .into_iter() on a borrowed ChunkedArray becomes .iter(). Same items (Option<T>), same semantics. 14 call sites:

• anndata/src/data/array/dataframe.rs (11)
• anndata/src/anndata/dataset.rs (1)
• pyanndata/src/anndata/backed.rs (1)
• pyanndata/src/anndata/dataset.rs (1)

Verified

cargo check --workspace                                  # clean
cargo test -p anndata -p anndata-hdf5 -p anndata-zarr    # 19 passed; 0 failed

rustc 1.97.1, x86_64-unknown-linux-gnu, Python 3.13.7, cmake 3.29.2.

I have not exercised the Python test suite — I could only get as far as compiling pyanndata and anndata_rs — so a look from someone who can run the bindings end-to-end would be worthwhile before merging.

Separately worth considering

Not included here, but it would retire this whole class of problem for downstreams:

pub use nalgebra_sparse;
pub use polars;

Consumers could then use anndata::polars::prelude::DataFrame and be structurally unable to resolve a mismatched version, instead of hand-shadowing this crate's choices and re-checking them each release. Both types already cross the public API boundary (DataFrame via set_obs/set_var, CsrMatrix via DynCsrMatrix). Happy to add it here or in a follow-up if you want it.