Integrate the multihit model into the DNSM framework

[willdumm]

This PR replaces #60, and is a companion to https://github.com/matsengrp/dnsm-experiments-1/pull/29

Here we add the multihit model to the DNSM. We chose to do so by storing the multihit model, as an object, as part of the DNSMDataset object, and bringing it into molevol.py.

This is a bit of a break from the previous work, which was able to stay in tensor land rather than object land. E.g. molevol.py now has objects from models.py.

Here is the rationale.

Previously, we could pre-compute the per-site rates and then all future calculations would be relative to those rates. To emphasize, the per-site rates don't depend on the branch lengths or anything.

Here, we have the multihit multipliers being applied after the codon level calculations. These codon calculations depend on the branch lengths and rates in a nonlinear way. Thus we can't precompute rates and then apply them in a straightforward way.

In theory we could stay more in tensor land by precomputing a 64*64 (this is codon by codon, because the hit classes depend on the source and dest codon) offset per site per sequence, and multiplying that after the codon aggregation step. However, that seems wasteful of memory and overkill.

So we have decided to apply the multihit correction using the forward function of the corresponding object, which is exactly what it's meant to do.

In any case, all of this work happens only in the branch length optimization phase of making the DNSMs, and not the transformer backprop.

[matsen]

Some small things here to fix.

[matsen]

Two last nitpicks.

[matsen]

Great!