feat(bf16): ml-dqn, ml-ppo, ml-supervised, ml-ensemble, ml-explainability compile clean
Dependency crates all compile with BF16:
- ml-dqn: noisy_layers, target_update, gpu_replay_buffer, branching → BF16
- ml-ppo: stubbed cuda_compile usage, PPO ops return errors (cold path)
- ml-supervised: liquid training host data → BF16 conversion
- ml-ensemble, ml-explainability: stubbed cuda_compile
Remaining: 108 errors in ml crate itself (Phase 3 Task 14 continuing).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
cacb1f8874