Add semantic SFT loss masks

[vivekkalyan]

Summary

Adds semantic SFT loss masks to ART trajectories and compiles them into the existing token-level label tensors during SFT preprocessing. This lets local/Megatron SFT train only selected assistant turns, such as the last GPT-4.1 continuation in a multi-turn production trace, while preserving the full prompt context.

Changes

• Add typed SFT loss mask specs on Trajectory: all, none, last_assistant, message_mask, and roles.
• Compile semantic masks in tokenizer preprocessing by rendering message-prefix spans and writing selected tokens into labels while keeping ignored positions at -100.
• Preserve loss_mask through JSONL SFT input, trajectory logging, and migration paths.
• Reject trajectory.loss_mask on serverless SFT for now, since this support is intended for local/Megatron SFT.
• Add Qwen disable-thinking response-part configs so empty <think> scaffolds are not trained as target content.

Why

Production-trace SFT needs to keep full multi-turn context but apply loss only to verified GPT-4.1 assistant outputs. Message-index masks in the dataset are fragile; this keeps the dataset representation semantic and lets ART compile it at tokenization time.

Validation

• uv run --no-sync pytest tests/unit/test_trajectory_parquet.py tests/unit/test_preprocessing_tokenize.py tests/unit/test_sft.py tests/unit/test_serverless_pipeline_trainer_compat.py -q
• uv run --no-sync prek run --all-files