fix: memory efficent backed anndata read

[nayib-jose-gloria]

Reason for Change

• anndata "backed" mode reads still loadX matrix into memory. Update read_h5ad to actually read backed objects
• feature_is_filtered validation check was previously loading X matrix into memory if bool column var.feature_is_filtered had any True entry. It would 1) subset the X matrix to only genes with featured_is_filtered is True and 2) validate (without chunking) that this filtered feature subsetted matrix had only zero values as declared by the feature_is_filtered column values.
• This would, as a result, load the X matrix into memory in order to 1) subset and then 2) count nonzero values and thus require higher resource provisioning in order to support use cases where the featured_is_filtered check was triggered.

Changes

• in feature_is_filtered validation check, leverage existing matrix nonzero counting function, _count_matrix_nonzero, which implements matrix chunking
  - pass in new optional arg, filter_by_column, to subset each matrix chunk by feature_is_filtered True cols before counting nonzeros