feat: MultiturnSFTDataset

[ysjprojects]

Summary

Adds MultiturnSFTDataset, a training-data counterpart to SFTDataset for
supervised finetuning on multi-turn conversations. Builds on the multi-turn
apply() support landed in #<PR number of 62ffd7b> (ChatML/Llama3/R1Base +
add_generation_prompt).

Motivation

litgpt's SFT pipeline (SFTDataset) only supports single-turn
instruction/output pairs — one prompt string in, one response string out,
with a single masked prefix span. There's no way to train on a full
conversation (multiple user/assistant turns) with per-turn loss masking:
every assistant turn should contribute to the loss while prior turns stay as
context, not get flattened into independent single-turn examples the way
deita.py/lima.py currently do.

What's included

• MultiturnSFTDataset (litgpt/data/base.py): takes a list of
  conversations (list[list[{"role", "content"}]]), and produces
  input_ids/labels with masking applied per-turn instead of a single
  prefix boundary.
  • Fast path when mask_prompt=False: one apply() + encode() call over
    the whole conversation.
  • When mask_prompt=True: since PromptStyle.apply() only returns a flat
    string with no turn-boundary info, turn boundaries are recovered by
    re-rendering growing prefixes of the conversation
    (apply(messages[:k], add_generation_prompt=False)) and diffing
    tokenized lengths — bos is applied consistently across every prefix so
    the diff stays aligned.
  • Requires prompt_style.supports_multiturn; raises a clear ValueError
    at construction otherwise.
  • Requires each conversation to end on an assistant turn (the training
    target); raises otherwise.
  • token_counts reports both raw (plain per-turn content, no template
    overhead) and raw_plus_prompt_template (actual sequence length),
    matching SFTDataset's existing convention used by the finetuning
    performance report.
• Exported from litgpt.data alongside SFTDataset.
• Tests in tests/data/test_base.py: parametrized coverage mirroring
  test_sft_dataset (mask_prompt / ignore_index / max_seq_length), a
  dedicated regression test for a longer alternating-turn conversation
  (catches a real bug found during development where inconsistent bos
  handling across prefixes corrupted turn boundaries), and both ValueError
  guard paths.

Out of scope (follow-up)

This PR is infrastructure only — no DataModule wires up
MultiturnSFTDataset yet. A follow-up PR will add a JSONConversations
(or similarly named) DataModule so litgpt finetune can actually consume
multi-turn JSON data end-to-end. Also out of scope: inference-side
multi-turn support (litgpt chat history, LLM.generate(messages=...)).

Test plan

• pytest tests/data/test_base.py tests/test_prompts.py — all passing.
• ruff check / ruff format --check clean.