Device-aware bf16 loading + T2V large/xlarge slice routing

Disable background warm (restart loop); lazy-load with adequate budget

Post-launch background warm of Wan 2.1 T2V-14B (COW for workers)

Cold-load-aware ZeroGPU duration budgets

Phase 1 fix — drop MoE to large + CPU offload (PRO compat)

Phase 1 — Wan Studio T2V/I2V on Wan 2.1 + 2.2 (commit fec10de)