Add SLIME per-token loss option

[Legendarylibr]

Summary

• add calculate_per_token_loss support to local SLIME config, manifest, trainer, and Forge UI
• store token-level rollout logprobs for opt-in token-weighted PPO loss while preserving default per-sample loss
• document the upstream-style option and add focused tests for config projection and loss normalization

Validation

• make ci-fast (1322 passed, 17 deselected, 1 existing Python 3.14 aiosqlite warning)
• .venv/bin/ruff check seiso/slime_single_gpu/config.py seiso/slime_single_gpu/trainer.py seiso/training/config.py tests/test_slime_single_gpu.py tests/test_training_config.py
• .venv/bin/python -m pytest tests/test_slime_single_gpu.py tests/test_training_config.py tests/test_multi_gpu.py -q
• cd forge-ui && npm run build