Options for dealing with missing flanking data in samples

[hyanwong]

As discussed with @jeromekelleher, it would be useful to be able to produce inferred tree sequences where the input samples have missing data to the left and right of a known sequence (e.g. fragments from a sequencer). This ability has been removed from #169, as it seems easier and more flexible to do this after inference-with-imputation on the fragmented sequences. For example, this would allow large sections of missing data in the middle of a sample (instead of only in the flanking regions) to be marked as "truly missing and not for imputation"

I intend to write a function to do this independently of the inference process. All this needs to do is to take an inferred TS and its corresponding SampleData file, remove the edges that link a sample to the tree at sites that are missing in that particular sample, and simplify(keep_unary=True). This issue replaces #153, and should eventually subsume #173.

[hyanwong]

I guess this is a function something like:

ts_with_missing = tsinfer.recreate_missing(sample_data, ts, flanking=True, min_num_sites=1, min_length=None)

Where min_num_sites =2 would mark as missing (i.e. remove edges from) any stretches in the sample_data file where 2 or more adjacent sites in a sample are marked as missing: this would allow the odd blip of missing data to be imputed, but remove imputed regions that span long lengths. Similarly, min_length=10 would mark as missing all sites where contiguous stretches of missing data in the sample file spanned more than 10 units of sequence length.

Running this on an inferred TS should allow us to round-trip data from a sample data file with missing data.