Adding per-base inference

[matsen]

• New family of models: RSXX for Rate Substitution models. These generate csps as well, which are conditional substitution probabilities.
• Generalizing everything to make that work

Note this that appears everywhere:

        # When we have an N, set all the CSP logits to 0, resulting in a uniform
        # prediction. There is nothing to predict here.
        csp_logits *= masks.unsqueeze(-1)
        # As described elsewhere, this makes the WT base have a probability of 0
        # after softmax.
        csp_logits += wt_base_modifier

[matsen]

Dropping this model shape:

            f"{prename}_cnn_med_orig": models.CNNModel(
                kmer_length=3,
                kernel_size=11,
                embedding_dim=9,
                filter_count=9,
                dropout_prob=0.1,
            ),

[matsen]

Here's a diagram of the RS CNNs:

[mmjohn]

I'm running into a few issues with unit tests on ermine in epam conda env.

First, I updated tests/test_framework.py by swapping create_mutation_and_base_indicators with encode_mut_pos_and_base (issues mentioned here). This test now passes.

Remaining errors:

FAILED tests/test_netam.py::test_write_output - RuntimeError: Parent directory _ignore does not exist.
FAILED tests/test_netam.py::test_crepe_roundtrip - RuntimeError: Parent directory _ignore does not exist.
ERROR tests/test_dnsm.py::test_crepe_roundtrip - FileNotFoundError: ``/Users/matsen/data/wyatt-10x-1p5m_pcp_2023-10-07.first100.shmple.hdf5`` does not ...

The first two can be solved by making the directory. It might be more useful to add a check for the directory and create if necessary as we do here in epam.

For the last issue, can this file live somewhere accessible?