Revert SE BF16 fallback — produced garbage output

The dequantize_nvfp4 path for shared expert made output WORSE (random
Chinese tokens, gibberish) vs NVFP4 GEMM which at least produces 'OK'.
The SE NVFP4 GEMM is working; the dequant scale computation was likely
wrong. Keeping BF16 router gate (which improved output from 'response'
loop to 'OK').

single_shot_inference.py

@@ -1359,9 +1359,7 @@ def main():
         se.set_fused_swiglu(True)
         # EAGERLY process shared expert weights
         se._ensure_initialized()
-        # BF16 fallback for shared expert — dequantize NVFP4 weights to BF16
         se._use_runtime_gsa = True
-        se.enable_bf16_fallback()  # sets _fused_swiglu=False, pre-materializes BF16 weights
         se_runners[li] = se
         if (li+1) % 10 == 0: print(f"  Built {li+1}/{n_layers} MoE layers")
         torch.cuda.empty_cache()