`--bias-AA-per-residue` doesn't work when `--model="proteinmpnn"`.

[Barry0121]

As title.

I specified the --bias-AA-per-residue pointing to a json file written same as the given example. Moreover, I tested this result with --model="ligandmpnn" and things works as expected.

However, when I'm using proteinmpnn weights, the outputs from the model don't respect the specified bias at each those specified positions. In addition, I've also tested the original ProteinMPNN implementation, and uses --bias_by_res_jsonl with a matrix to specifiy the AA bias at each residue position, and I got a similar phenomenon, i.e. the output doesn't follow the given bias assignment, but rather falling back to pure random assignment.

Any idea why or how to fix?