sparse.COO support for harmonize_missing_values and infer_feature_types

[sueoglu]

ehrdata can store a numeric .X/layer as sparse.COO tensor, but two core functions don't handle it correctly::

• infer_feature_types crashes on sparse.COO input:

>>> ed.infer_feature_types(edata)  # edata.X is sparse.COO
NotImplementedError: The `order` parameter is not supported

It tries to reshape the sparse array the same way it reshapes a dense one, which sparse.COO.reshape doesn't support. Since several downstream functions (e.g .ep.pp.scale_norm) call infer_feature_types automatically whenever edata.var["feature_type"] isn't set yet, this blocks normal usage of any sparse.COO-backed EHRData

• harmonize_missing_values silently does nothing for sparse.COO:
  For a sparse array, the implicit 0 fill value is ambiguous, it can mean "not measured" or "measured as zero", but harmonize_missing_values currently treats all numeric layers (dense or sparse) as a no-op, so there's no way to mark genuinely-missing sparse entries as NaN.

TODO

• infer_feature_types should work on a sparse.COO X/layer without densifying the whole array
• harmonize_missing_values treats a sparse.COO array's implicit 0 as missing by default, without densifying, with an opt-out mechanism for columns whose 0 is a real, measured value rather than a missing one
• No behavior change for other types