Fill holes in Genotype encoding API

[droazen]

The Genotype::encode_genotype() API currently doesn't allow you to phase only some alleles and not others, and also has no function suitable for encoding multiple genotypes at once with phasing.

Note that if we end up removing the requirement to manually encode genotypes as proposed in #371, we will still need to provide the equivalent of the above functionality in VariantBuilder.

[MauricioCarneiro]

We need @vruano and @rpoplin in this discussion. Imagine they were writing a variant caller that is phasing aware. They will need to write these out in a sensible way. There is no good model API neither in Tribble or htslib for that. We need to come up with one and people will follow.

[droazen]

One possibility is to have a simple GenotypeAllele struct -- eg.,

struct GenotypeAllele {
    int32_t allele;
    bool phased;
};

Then clients could pass a vector to the builder instead of a vector<int32_t>. This would remove the need to specify phasing information separately, but the disadvantage is that the user's genotype vector could no longer be encoded in-place and passed to htslib -- we would always have to make a copy.

[droazen]

Spoke with Ryan and he feels that it would be very rare for a tool to need to create genotypes with mixed phasing like 0/1|2, so we definitely shouldn't slow down the common use cases for the sake of allowing this.

One possibility is to continue to have functions that take vector<int32_t> for the unphased case, and if phasing is desired you would either create a vector as proposed above, or pass in a companion vector with the phasing information.

[MauricioCarneiro]

I like what @rpoplin said. So anything that requires crazy phasing we can
just do inefficiently. Until it becomes a use case for us. But keep the
common case fast.