Tighten ZeroGPU durations to p95-aligned values (10s fast, 30s high-detail)

Based on 410 samples: fast p95=5s (max 13.3s), high-detail p95=18.3s
with 2.8% exceeding 30s. Accepts ~5% failure rate; users can retry
with remaining daily quota.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

app.py

@@ -352,7 +352,7 @@ def infer(images_b64_json, prompt, seed, randomize_seed, guidance_scale, steps,
         raise
 
 
-@spaces.GPU(duration=lambda *a, **kw: 30 if (len(a) > 7 and a[7] == "fast") else 90)
+@spaces.GPU(duration=lambda *a, **kw: 10 if (len(a) > 7 and a[7] == "fast") else 30)
 def _infer_gpu(pil_images, prompt, seed, guidance_scale, steps, width, height, mode):
     _cuda_ok = torch.cuda.is_available()
     timer = _InferTimer(_cuda_ok)