Add baseline and depth recurrence submissions (1xH100 20min runs)

[henrycashe26]

Two submissions, both trained on 1xH100 for 20 minutes (competition standard is 8xH100 for 10 min, so these are underpowered runs).

Baseline 10L Int5-MLP hit 1.2604 val_bpb. This was the #1 leaderboard config (as of a few days ago) reproduced on less compute. 10 layers, mixed int5/int6 QAT, BigramHash 10240, SWA, Muon optimizer. 25.5M params, 15.8MB artifact. Got through ~1700 steps at 695ms/step.

Depth Recurrence 4Lx3Loop hit 1.3752 val_bpb. 4 unique transformer layers shared across 3 loop iterations, with LoRA rank-32 per loop and learned level signals. 11.6M params, 8.3MB artifact (7.7MB headroom left under the 16MB cap). ~1417 steps at 847ms/step. Quantization roundtrip cost 0.047 bpb because shared weights serve different input distributions at each loop, which doesn't quantize cleanly.

Notes
Trained on 1xH100 SXM, not 8xH100, so step counts are well below a standard run. Single seed (42) for both, no multi-seed significance claims. Loss was still dropping for both models when we cut them off.