perf: fix low-hanging performance issues in MetaCAT and linking

[bgriffen]

Summary

Four small, safe performance fixes across the MetaCAT and linking components:

• Scope batch padding to current batch (ml_utils.py) — create_batch_piped_data was computing max_seq_len over the entire dataset on every batch call, then slicing data[start_ind:end_ind] three times. Now scoped to a single batch slice. Reduces unnecessary padding and iteration.

• In-place similarity update (vector_context_model.py) — _preprocess_disamb_similarities was copying the similarities list, clearing it, and rebuilding via list comprehension. Replaced with a simple in-place loop, eliminating three intermediate allocations in the disambiguation hot path.

• O(1) dict lookup instead of O(n) values scan (data_utils.py) — undersample_data and encode_category_values checked membership against category_value2id.values() (linear scan) every iteration. Since label_data dicts use the same keys, check against the dict directly.

• Append instead of list concatenation (ml_utils.py) — _eval_predictions used dict.get(k, []) + [item] which allocates a new list per iteration. Switched to setdefault + append for O(1) amortized insertions.

All changes are behaviour-preserving — no new dependencies, no config changes, no API changes.

Test plan

• Existing medcat-v2 unit tests pass across Python 3.10–3.13
• MetaCAT training + inference produces identical results
• Linking disambiguation produces identical concept rankings

[mart-r]

Thanks for the contribution!

Looks good to go. Minor changes, but clear improvement.