fix(template): truncate token type IDs with retained positions

[taking-lying-flat]

PR type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

PR information

When an encoded multimodal sample exceeds max_length, Template._truncate() truncates input_ids, labels, loss scales, and mm_token_type_ids but leaves token_type_ids at its original length. For example, left-truncating a 12-token Gemma3-Vision sample to 8 tokens leaves 12 token types and moves the image tokens away from their original positions. The collator pads these fields independently, preserving the mismatch.

Truncate token_type_ids with the same retained-position mask as the other per-token fields. This also handles non-contiguous retained positions caused by protected image placeholders. The production change is three lines in the shared template implementation.

Extend the existing Gemma3 template tests with left/right truncation cases containing text around one or two images. Check the exact values of all five per-token fields and their shapes and values after collation, including non-contiguous retention cases.

Experiment results

• The four new truncation cases fail on the original implementation with misaligned token types.
• python -m pytest tests/general/test_gemma3_template.py tests/general/test_moss_vl.py -q: 14 tests and 4 subtests passed.
• pre-commit run --files swift/template/base.py tests/general/test_gemma3_template.py: all applicable hooks passed.
• git diff --check: passed.

Validation uses synthetic encoded samples through the real truncation and collator paths. Full-model forward and training were not run.