Fixing BS=1 handling

[jannisborn]

No description provided.

[kwehden]

The core fix looks correct. Argument-less `torch.squeeze()` drops every singleton dim, so with BS=1 the input `[1, seq_len, 1]` collapses to `[seq_len]` and the batch dimension is lost. `.squeeze(-1)` only removes the trailing dim, giving `[1, seq_len]` and preserving the batch — this is the right fix and is also safer in general.

Two follow-ups:

1. Same bug remains in `scripts/test_toxsmi.py:203`:
   ```python
   yhat, pred_dict = model(torch.squeeze(x.to(device)))
   ```
   This is the identical pattern, so inference will still break at BS=1. Suggest applying the same `.squeeze(-1)` change here so BS=1 is fixed end-to-end.

2. .gitignore nit (optional): the file is named `.DS_Store`, so the conventional entry is `.DS_Store` (or `**/.DS_Store`) rather than `*.DS_Store`; also missing a trailing newline.

LGTM to merge once the test script line is addressed.