Add executor affinity for numpy/torch thread safety

Workers are now assigned a fixed executor_id at creation. All calls
from the same worker go to the same executor thread, preventing
thread state corruption in libraries like numpy and PyTorch that
have thread-local state.

This fixes segfaults when using sentence-transformers or other
ML libraries that depend on consistent thread affinity.

Author: benoitc

c_src/py_exec.c

@@ -791,8 +791,16 @@ static int executor_enqueue(py_request_t *req) {
 
         case PY_MODE_MULTI_EXECUTOR:
             if (atomic_load(&g_multi_executor_initialized)) {
-                /* Route to multi-executor pool */
-                int exec_id = select_executor();
+                /* Route to multi-executor pool.
+                 * Use worker's assigned executor for thread affinity if available.
+                 * This ensures libraries like numpy/torch that have thread-local
+                 * state always run on the same thread for a given worker. */
+                int exec_id;
+                if (req->worker != NULL && req->worker->executor_id >= 0) {
+                    exec_id = req->worker->executor_id % g_num_executors;
+                } else {
+                    exec_id = select_executor();
+                }
                 multi_executor_enqueue(exec_id, req);
                 return 0;
             }

c_src/py_nif.c

@@ -1418,6 +1418,11 @@ static ERL_NIF_TERM nif_worker_new(ErlNifEnv *env, int argc, const ERL_NIF_TERM
     worker->has_callback_handler = false;
     worker->callback_env = NULL;
 
+    /* Assign executor affinity for thread-safe library support (numpy, torch).
+     * Each worker gets a fixed executor to ensure all calls from the same
+     * worker go to the same thread, preventing thread state corruption. */
+    worker->executor_id = select_executor();
+
     PyGILState_Release(gstate);
 
     ERL_NIF_TERM result = enif_make_resource(env, worker);

c_src/py_nif.h

@@ -390,6 +390,9 @@ typedef struct {
 
     /** @brief Environment for building callback messages */
     ErlNifEnv *callback_env;
+
+    /** @brief Assigned executor ID for thread affinity (-1 = round-robin) */
+    int executor_id;
 } py_worker_t;
 
 /* async_pending_t and py_async_worker_t removed - async workers replaced by event loop model */