sgkit: Support call_genotype_mask?

[benjeffery]

Often when running tsinfer, a subset of the variants are wanted for inference. Making a subset sgkit dataset isn't ideal, as it means making a copy of the dataset and rechunking (see https://github.com/pystatgen/sgkit/issues/981).

One option is to have tsinfer support a mask as part of the input dataset. I think I'm right in saying that for tsinfer variant_mask would make the most sense - but in sgkit call_genotype_mask is standard. We could use call_genotype_mask and error if a variant has a mixture of masked and unmasked calls?

The main changes to enable this are in the chunk iterator, which would need to skip the masked rows.

[jeromekelleher]

I think we need to support masks, because we want to support incremental exploration of the data with extra levels of filtering, without needing to store copies.

[benjeffery]

After some discussion it seems best to add a variant_mask rather than use call_genotype mask. Also to use that array if present, rather than have the user specify on each call to tsinfer.