SFT for multi-turn tool call data is not supported properly

SFT Tool Support Proposal

Problem

1. When --apply-chat-template is used, the message history (list[dict]) gets converted to a string in data.py. This breaks gen_multi_turn_loss_mask_qwen and other functions in mask_utils.py, which expect a list[dict].

2. --tool-key requires --apply-chat-template, which triggers the same string conversion issue.

3. In mask_utils.py, apply_chat_template is called on each message individually without the tools parameter, so tool-related messages (assistant with tool_calls, tool role responses) are not formatted correctly.

Downstream Effects

• Without these flags: Tool definitions are never added to the conversation, so the model never sees them during training.
• With these flags: A fatal error is thrown because mask_utils receives a string instead of list[dict].
• Even if it didn't crash: apply_chat_template without tools= produces incorrect tokenization for tool-calling messages.

Proposal

A non-invasive fix with two changes:

1. data.py - Store tools in metadata

When --tool-key is provided:

• Extract tools from the data
• Store them in sample.metadata["tools"]
• Keep messages as list[dict] (do NOT apply chat template)

2. mask_utils.py - Pass tools to apply_chat_template

Update the mask generation functions to:

• Accept an optional tools parameter
• Pass tools to apply_chat_template so tool-calling messages are formatted correctly

Also this fixes Issue #289

Read data.py and mask_utils.py, starting with the --apply-chat-template and --tool-key handling and the gen_multi_turn_loss_mask_qwen path. Trace how message history and metadata reach mask generation. Done means tool definitions remain available through metadata, messages stay as list[dict], and apply_chat_template receives tools for tool-calling messages without the reported failure.