fix(wasm): remove prefill sleep — restores token streaming

quant.h

@@ -15462,14 +15462,13 @@ int tq_generate(tq_model_t* model, tq_tokenizer_t* tokenizer,
     }
 
     /* Prefill: process all prompt tokens.
-     * On Emscripten with ASYNCIFY, yield every 2 tokens so the browser
-     * can repaint (shows "Thinking..." and avoids "page unresponsive"). */
+     * NOTE: No emscripten_sleep() here — the call stack during tq_forward()
+     * is too deep for ASYNCIFY to unwind (matmul → SIMD kernels). Adding
+     * sleep here breaks ASYNCIFY for the entire generate call, including
+     * the token streaming callback. The browser shows "Thinking..." via
+     * requestAnimationFrame before entering this blocking prefill. */
     for (int i = 0; i < n_prompt; i++) {
         tq_forward(model, state, prompt_tokens[i], i);
-#ifdef __EMSCRIPTEN__
-        extern void emscripten_sleep(unsigned int ms);
-        if (i % 2 == 1) emscripten_sleep(0);
-#endif
     }
 
     /* Repetition penalty setup */