feat(gkd): on-policy distillation wiring + spec (armed)

[ronaldtse]

GKD rung implemented per its EXPERIMENTS.md registration (armed in #147):

• src/gpu/gkd.py — modal-free pure-torch helpers: gkd_weight (full ratio to 2/3 of training, linear to zero over the final third), token_logprobs, reverse_kl (student-only gradients; teacher scored under no_grad). 7 unit tests (importorskip torch, the MTP-test pattern).
• distill_sequence wiring — teacher moved back to cuda for scoring; every sample_every=4 steps a sub-batch of 2 student-sampled sequences (temp 1.0, cap 1024) is scored by both models on (prompt -> sample) and the reverse-KL enters the loss at the annealed weight. Control-verbatim otherwise: same units, schedule, Muon, 30k units.
• Spec ara-diac-small-2-gkd (run-011-r7-muon-gkd) copied field-for-field from run-006 with only the gkd block added.
• Fixed pre-existing B905 in scripts/figures/frontier.py (was making main red on ruff).

Launch remains gated on a GPU slot (G2b + one other app currently hold the budget; GKD launches when a slot frees).