Fix aliased list references in _tokenize_fn

[Mr-Neutr0n]

Summary

• input_ids = labels = [...] in _tokenize_fn assigns both variables to the same list object, so mutating one silently mutates the other. The same applies to input_ids_lens = labels_lens = [...].
• The current call-site in preprocess() masks this with a copy.deepcopy, but the bug is latent: any future caller that reads labels from the returned dict will get the exact same object as input_ids, leading to silent data corruption when the label-masking step overwrites source tokens with IGNORE_INDEX.
• This PR builds separate lists for input_ids/labels and input_ids_lens/labels_lens so each is an independent object.

Repro

# Demonstrates the aliasing problem:
a = b = [1, 2, 3]
a[0] = 99
print(b)  # [99, 2, 3] — b was mutated too

The same thing happens in _tokenize_fn with input_ids and labels.

Test plan

• Verified that the fix produces correct, independent lists
• No functional changes to the training pipeline (only removes a latent aliasing hazard)