Add sparse.COO support to harmonize_missing_values and infer_feature_types

[sueoglu]

closes #305

Both functions previously either crashed or silently no-op'd on a sparse.COO X/layer. This adds proper support to both, without ever densifying the array.

infer_feature_types

• sparse.COO is always numeric under the binsparse spec this project follows, so classification per variable only needs to decide "is this column exactly {0, 1}?" which is answerable directly from X.coords/X.data/X.fill_value
• _detect_feature_types_sparse_coo helper implements this; infer_feature_types branches to it for sparse.COO input and keeps the existing dataframe-based path unchanged for everything else.

harmonize_missing_values

• For sparse.COO, the implicit 0 fill value is now treated as missing (swapped to NaN) by default (done by flipping the array's fill_value from 0 to NaN, so the array stays sparse)
• vars argument: variable names whose 0 is a real measured value, not missing. Previously-implicit zeros of these columns are materialized as explicit stored 0 entries (so they dont get changed to the new NaN fill value); every other column stays sparse
• Left untouched: boolean arrays, scipy.sparse

Tests: added coverage in test_feature_types.py for 2D/3D sparse.COO harmonization, vars exclusion, an unknown-var error, scipy.sparse being unaffected, and sparse.COO feature-type inference (including binary detection and the all-NaN-column error). Two pre-existing IO round-trip tests (test_h5ed.py/test_zarr.py) updated to pass harmonize_missing_values=False, since they test binsparse encoding fidelity specifically and predate this behavior change.

[eroell]

Please add a benchmark with compute time and memory consumption in the PR comment, where a large dense 3D and a large sparse 3D, and also the 2D sparse and dense arrays are tested.

The sparse COO version should run for a ~500k x 1000k x 90 tensor of about 1% data density.

[sueoglu]

Benchmark with compute time and memory usage

Used time.perf_counter() for wall-clock time, tracemalloc for peak Python-allocator memory. I wasnt able to login to cluster and couldnt run it with a large-scale case representing EHR dimensions (~500k × 1000k × 90).

case	shape	build (s)	build peak (MB)	infer_feature_types (s)	infer peak (MB)	harmonize_missing_values (s)	harmonize peak (MB)	nnz
dense_2d	2000×3000	0.09	72.0	30.75	6.2	0.00	0.0	6,000,000
sparse_2d	2000×3000 @1%	0.02	4.6	0.03	0.4	0.00	1.6	59,712
dense_3d	500×800×90	0.41	432.0	181.75	150.1	0.00	0.0	36,000,000
sparse_3d	500×800×90 @1%	0.10	36.3	0.03	1.4	0.01	9.3	358,173

• infer_feature_types on dense 3D takes ~182s vs ~0.03s for sparse.COO at identical shape