forceWasm init-needs-wasm respawn

Chose one fresh worker with sticky WorkerInitData.forceWasm=true over switching the existing worker because onnxruntime-node caches the module after reading __LORE_ORT_BINDING_PATH__; only a new module graph can recover from native model-load incompatibility. Reject/drop the pending request through canonical shouldPostPerRequestError handling before resubmission, ignore stale worker error/exit events, preserve forceWasm across OOM respawns, and treat a second init-needs-wasm as terminal initialization failure to prevent loops.