Fix step-done log ordering vs text_encoder offload/pre-VAE-decode

The step-done print used a timestamp captured at callback entry but
printed after the offload/pre-VAE-decode logging (which use fresh
perf_counter() calls), making log lines appear out of chronological
order on the final step. Print step-done immediately after computing
it instead.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

app.py

@@ -575,6 +575,9 @@ def _make_step_callback(steps, timer, t0, mode: Mode, cuda_ok: bool = False):
         if step_idx == 0:
             timer.mark("first_step")
         timer.mark("last_step")  # overwritten each step; final value = end of last step
+        delta_ms = (now - (step_times[-2] if len(step_times) > 1 else t0)) * 1000
+        tag = " ← includes cold-start (offload hook install + first weight transfer)" if step_idx == 0 else ""
+        print(f"[infer] step {step_idx+1}/{steps} done — {delta_ms:.0f}ms{tag} | t={now-t0:.1f}s")
         # Text encoder is done after prompt encoding, before the denoising loop starts.
         # Dropping it (~15GB) ahead of VAE decode's fp32-upcast memory spike only pays off
         # when that spike is big enough to need the headroom (see Mode.offloads_text_encoder_before_decode).
@@ -596,9 +599,6 @@ def _make_step_callback(steps, timer, t0, mode: Mode, cuda_ok: bool = False):
         if step_idx == steps - 1 and cuda_ok:
             print(f"[infer] pre-VAE-decode — {_gpu_mem_str(True, sync=True)} | t={time.perf_counter()-t0:.1f}s")
             torch.cuda.reset_peak_memory_stats()
-        delta_ms = (now - (step_times[-2] if len(step_times) > 1 else t0)) * 1000
-        tag = " ← includes cold-start (offload hook install + first weight transfer)" if step_idx == 0 else ""
-        print(f"[infer] step {step_idx+1}/{steps} done — {delta_ms:.0f}ms{tag} | t={now-t0:.1f}s")
         return cb_kwargs
     return _step_cb