fix: descriptive tqdm labels — uint8→NVFP4 and NVFP4→FP8/BF16

Makes it crystal clear what's happening:
- Experts: direct uint8→float4 view-cast (Blackwell native, no BF16)
- Convert: NVFP4→FP8/BF16 for attention weights (non-expert path)

vllm/patches/deepseek_v4.py

@@ -429,7 +429,7 @@ class DeepseekV4MegaMoEExperts(nn.Module):
         l2_fp4, l2_sf, l2_gs = [], [], []
 
         from tqdm import tqdm
-        for e in tqdm(range(self.num_local_experts), desc="  NVFP4 experts", unit="exp"):
+        for e in tqdm(range(self.num_local_experts), desc="  uint8→NVFP4 experts", unit="exp"):
             # ── L1: gate + up (fused) ──
             gate_w = self.w13_weight.data[e, :self.intermediate_size]  # (intermediate, hidden//2) uint8
             up_w = self.w13_weight.data[e, self.intermediate_size:]    # (intermediate, hidden//2) uint8
@@ -1623,7 +1623,7 @@ class DeepseekV4Model(nn.Module):
         _shard_index = self._build_shard_index("/model") if os.path.isdir("/model") else None
 
         from tqdm import tqdm
-        for layer_idx, layer in tqdm(enumerate(self.layers), total=len(self.layers), desc="  NVFP4 convert", unit="layer"):
+        for layer_idx, layer in tqdm(enumerate(self.layers), total=len(self.layers), desc="  NVFP4→FP8/BF16 convert", unit="layer"):
             attn = layer.attn
             
             # FP8 conversion: only wo_a