cleanup: remove debug prints, ready for testing

Current state:
- Token indices on CPU (avoids CuTeDSL GPU memory corruption)
- Scale assembly uses per-expert swizzle + scatter (matches reference)
- compute_activation_global_scales warmup gets ~0.97 cosine
- expert_offsets passed without leading 0 (matches pipeline)
- layertest + cudagraph_test pass

vllm/nvfp4_cutedsl.py

@@ -242,12 +242,6 @@ class CuTeDSLMoERunner:
             sorted_token_ids = token_indices[sort_idx.cpu()].to(device)
             slot_hidden = hidden_states_sample[sorted_token_ids]
             
-            # Debug: verify slot_hidden
-            torch.cuda.synchronize()
-            _slot_check = sorted_token_ids[:8].cpu().tolist()
-            _slot_amax = slot_hidden.abs().max().item()
-            print(f"  Warmup: sorted_token_ids[:8]={_slot_check}, slot_hidden amax={_slot_amax:.6f}")
-            
             # L1: get exact gs from quantize_to_nvfp4
             _, _, l1_gs = quantize_to_nvfp4(slot_hidden)