…ForCompletionOnlyLM (huggingface#3223)

Refactors the masking logic in `DataCollatorForCompletionOnlyLM` to correctly handle conversations with multiple instruction roles (e.g., user, tool) and consecutive assistant turns, enabling its use for more complex dialogue formats like agent trajectories.

Previously, the collator assumed a strict alternation of a single instruction template and a response template (e.g., User -> Assistant). This failed for:
1.  Datasets with multiple instruction roles (e.g., user prompts and tool calls).
2.  Sequences with consecutive assistant messages (e.g., Assistant -> Assistant).

This commit addresses these limitations:
- Updates `__init__` to accept a list of strings or pre-tokenized IDs for `instruction_template`, allowing multiple distinct instruction roles.
- Rewrites the core masking logic in `torch_call`:
    - It now identifies all occurrences of response and all specified instruction templates.
    - For each assistant response, it unmasks tokens from the end of its template up to the beginning of the *next* instruction template or the sequence end.
    - Correctly handles consecutive assistant turns by masking the template tokens of subsequent responses while unmasking their content.
- Adds comprehensive unit tests (`test_masking_*`) covering multi-role scenarios, consecutive assistant messages, left-padding, and initialization with tokenized templates.

This allows `DataCollatorForCompletionOnlyLM` to process conversational data commonly found in ChatML formats and agent fine-tuning datasets.

Related: huggingface#1994, huggingface#2545