fix: align InstaNovo tests with the default beam token log-probabilities prefix

## Summary

`tests/datasets/data_loaders/test_instanovo.py` uses:

```python
"token_log_probabilities": "predictions_token_log_probabilities_"
```

The shipped config and real InstaNovo v1.2.2 output use:

```python
"token_log_probabilities": "predictions_token_log_probabilities_beam_"
```

The tests are internally consistent, so they pass without exercising the actual default convention.

## Change

- Update the standard loader fixture and its test data to use
`predictions_token_log_probabilities_beam_`.
- Add an assertion for the three prefixes shipped in
`winnow/configs/data_loader/instanovo.yaml`.
- Leave tests for custom beam-column prefixes unchanged.

This is test alignment only. No loader behaviour, schema, or defaults should change.

Start with tests/datasets/data_loaders/test_instanovo.py and compare its standard loader fixture and test data with the three prefixes in winnow/configs/data_loader/instanovo.yaml. Run the InstaNovo test file; done means the standard data uses predictions_token_log_probabilities_beam_, all three shipped prefixes are asserted, and custom beam-column prefix tests remain unchanged.

Written by the indexing model from the issue text.