fix(spaces): switch training Space to A10G Small, tune notebook for 24 GB

fix(notebook): Cell 3 load_dataset per split + write JSON (was snapshot_download finding 0 .json files)

feat(training): A10G-optimised pipeline — auto train.py, Dockerfile.train, GH Action sync

feat(training): Phase C6 — ablations, training curves, readme finalization

feat(training): Phase C4 — GRPO training, SFT warmstart, rollout, custom trainer