Abolish unfilled likelihoods and revamp VariantAnnotator

[davidbenjamin]

Here's the rundown of this PR:

We had this awkward implementation of AlleleLikelihoods called UnfilledReadLikelihoods, which didn't contain likelihoods but did have pileups. This was necessary because the annotation engine expects likelihoods but VariantAnnotator only has the reads. Several annotations had fallback code to annotate a likelihoods object that had no likelihoods. @vruano This is the class that you disliked so much in your recent code review of #5783

A few issues with this state of things:

• Torturing the definition of AlleleLikelihoods, which forced the class to have methods like hasLikelihoods().
• VariantAnnotator only applied the few annotations that had custom pileup-based fallback code.
• Lots more annotation code for the fallback mode.

So the first step was the option that @lbergelson and @jamesemery liked most: create a regular likelihoods object in VariantAnnotator by hard-assigning of each read to the allele it best matches. This is exactly what all the custom fallback modes were doing in effect, but now it's implemented in one place instead of six or so. This lets us delete UnfilledLikelihoods and also lets VariantAnnotator apply any annotation.

@ldgauthier Since the most non-trivial aspect is the new integration test I'm inclined to assign you the review, but a case could be made for someone on the engine team.

This completely broke the VariantAnnotator tests, which were based on exact matches. This had been an issue before and has always been a bit of a nuisance, but now overhauling the tests became completely unavoidable. So, I rewrote all the tests and wrote a rigorous test based on concordance with annotations from Mutect2.

If I were reviewing I would start with the new code in VariantAnnotator that constructs the likelihoods object from the reads and verify that it is just a more polished version of the fallback code that several annotations used to have. Then I would look at the new VariantAnnotator integration tests. Some of the tolerances are fairly liberal but it's worth noting that much of the old exact match "truth" annotations were completely bogus. This is better than what we had before by a long shot but it's still use at your own risk.

[ldgauthier]

Some questions about preserving old test behavior.

[ldgauthier]

Are we supposed to match alleles with dbSNP now? There certainly was a time when we didn't, i.e. a different ALT at the same position would still get annotated.

[davidbenjamin]

That's unchanged. The dbSNP annotation still cares only about position.

[davidbenjamin]

Back to @ldgauthier

[ldgauthier]

LGTM!