Vendored generate() (non-streaming + streaming) gains an optional framework-agnostic progress_callback(current, total) fired once per AR loop step. modules/generation.py wraps it with comfy.utils.ProgressBar (interrupt check per step, dynamic total via update_absolute, guaranteed final 100%% in finally). ASR reports per-token progress via an HF BaseStreamer (greedy/sampling only; beam search falls back to 0->100%%). 25 new tests; full regression 560p/5f(pre-existing)/4s, zero new failures.
