to_zarr then from_zarr drops the phased flag, so a saved simulation comes back unphased #403

Since #401, HaplotypeMatrix.from_ts gives phased=True, but a matrix saved with to_zarr and read back with from_zarr has phased=False. Phase-dependent statistics on it then emit UnphasedDataWarning, although the data are the same phased haplotypes. A common workflow (simulate with msprime, save to VCZ, analyze later) hits this.

On main (53ea50b):

import tempfile, warnings
import msprime
from pg_gpu import HaplotypeMatrix, selection

ts = msprime.sim_mutations(
    msprime.sim_ancestry(10, sequence_length=1e5, recombination_rate=1e-8,
                         population_size=1e4, random_seed=1),
    rate=1e-8, random_seed=1)
hm = HaplotypeMatrix.from_ts(ts)
print(hm.phased)                                   # True
path = tempfile.mkdtemp() + "/sim.vcz"
hm.to_zarr(path, format="vcz", contig_name="1")
back = HaplotypeMatrix.from_zarr(path)
print(back.phased)                                 # False
with warnings.catch_warnings(record=True) as w:
    warnings.simplefilter("always")
    selection.garud_h(back)
print([type(x.message).__name__ for x in w])       # ['UnphasedDataWarning']

Cause

• The VCZ writer (zarr_io.write_vcz, called by to_zarr) does not write call_genotype_phased, the VCZ field that records phase per call.
• from_zarr does not read it. Its phased argument defaults to False.

Proposed fix

1. to_zarr writes call_genotype_phased, filled with the matrix's phased value (all True or all False). bio2zarr writes the same field from a VCF's | and /.

2. On load, use the field. Two options:

   • a. from_zarr(phased=None) by default: None means phased only when the store has call_genotype_phased and every call is phased; an explicit True or False still wins.
   • b. Keep phased=False as the default, and read the field only when the user passes phased='auto'.

   Option a keeps the round trip without user action. It also makes a VCZ store converted from a fully phased VCF load as phased, which goes a step past Phased attribute to HaplotypeMatrix and warnings in phase-dependent statistics. #401's rule that loaders do not inspect the input's phase (see the discussion on pg-gpu needs to know if HaplotypeMatrix is phased or not #168: partial phasing, for example from GATK, makes inference risky). Requiring every call to be phased keeps a partly phased store unphased. Option b keeps Phased attribute to HaplotypeMatrix and warnings in phase-dependent statistics. #401's rule and needs the user to ask for the store's phase.

3. A test: from_ts, then to_zarr, then from_zarr (eager and streaming) keeps phased=True, and the same round trip keeps phased=False for an unphased matrix.

GenotypeMatrix.to_zarr has no flag to write, so it is not affected.

Related: #168, #401.