feat: add reproducible reasoning SFT on LUMI

[BirgerMoell]

Summary

Adds the missing shared infrastructure needed to express a reproducible reasoning-SFT continuation on LUMI:

• supports a full accelerate.config_file launch profile and freezes it beside the generated SLURM job, enabling the tested multi-node FSDP shape without a bespoke launcher;
• adds a LUMI FSDP profile and a production-oriented 9B/16K reasoning-SFT config for the pinned Anneal-300B instruction checkpoint;
• adds dataset revision pinning through loading, offline prefetching, and data inspection;
• includes revision, subset, split, transform, and other row-defining fields in automatic run identity, including single-dataset runs;
• documents the assistant-only masking, tokenize-only gate, checkpoint/evaluation workflow, and the contract for an immutable token-balanced reasoning artifact.

The guide deliberately does not present row-weighted source mixing as equivalent to the current production recipe. Token allocation, cross-source deduplication, loop rejection, decontamination, and manifest creation remain explicit upstream data-curation responsibilities; this trainer consumes their materialized Parquet artifact.

Type of change

• Bug fix
• New feature
• Refactor
• Performance
• Documentation
• Maintenance

Validation

• Full test suite: 160 passed
• Ruff 0.9.10 lint and format checks passed
• Black check passed
• Python compilation and git diff --check passed
• Rendered the reasoning profile end to end and verified that the FSDP config is frozen into the run directory, forwarded with --config_file, and no DeepSpeed flags are emitted
• Verified that the reference token budget resolves to exactly 500 optimizer steps (524,288,000 packed tokens at global batch 64 and sequence length 16,384)