[datasets] Add Hermes trace support to the SFT pipeline

[taivu1998]

Add conversation transform hooks for dataset-specific message normalization and stable row IDs, then register the Hermes glm-5.1 and kimi trace splits for SFT. Adds regression fixtures/tests and a trace-focused pilot experiment.

[taivu1998]

🤖 Spec

Problem
Marin's conversation transform path only accepted raw adapter output, so datasets that need dataset-specific message cleanup or source-derived stable IDs could not be integrated without bespoke pipelines. Hermes traces need tool responses unwrapped into Marin's structured tool-message shape while preserving assistant <think> and <tool_call> content, and the dataset rows should keep source IDs.

Approach
Add optional message_postprocess_fn and row_id_fn hooks to TransformAdapter, with signature generation that omits unset hook fields so existing dataset output paths stay stable. Implement Hermes normalization in trace_normalization.py, register the glm-5.1 and kimi splits in instruction_datasets.py, and add a pilot SFT experiment plus focused fixtures and regression tests.

Key code

• transform_row(...) now applies adapter postprocessing before replacements/tool normalization and prefers row_id_fn when present.
• normalize_hermes_trace_messages(...) only rewrites well-formed <tool_response> tool turns; malformed payloads fall back to raw source content unchanged.
• hermes_trace_row_id(...) preserves dataset IDs and falls back to a content hash only when the source row has no usable id.

Tests
./infra/pre-commit.py --all-files --fix
UV_CACHE_DIR=/tmp/uv-cache-pr1 uv run --all-packages pytest tests/transform/test_agent_trace_conversation.py tests/transform/test_conversation.py tests/test_marin_chat_template.py -q
UV_CACHE_DIR=/tmp/uv-cache-pr1 uv run --all-packages python -c "import experiments.exp_hermes_trace_sft_pilot"