Isolate CPU fallback in spawn subprocess to prevent CUDA state poisoning

Eliminate in-decorator CPU fallback to prevent CUDA state poisoning

Stop CUDA state poisoning by not invalidating model caches during quota fallback

Add debug logging

Park AoTI cache and detach VAD tensors to prevent CUDA state poisoning

Prevent CPU fallback from poisoning CUDA state via GPU model parking

Fix GPU cross-user contamination: scope errors per-request, not process-wide

Eliminate CUDA state poisoning at root cause, not just detection

Fix CUDA state poisoning by eliminating CUDA ops outside GPU leases

Prevent CUDA state poisoning and add thread-safe GPU lifecycle management

Harden GPU fallback to prevent CUDA init from permanently poisoning process

Fix race condition where concurrent quota exhaustion permanently breaks GPU

Restructure src/ for cleaner modularity

Initial commit