Add resume capability to dequant script (skip already-done shards)

Verified our FP4 dequant is byte-identical to official transformers
MXFP4 implementation. Max diff = 0.0 across all values.

scripts/dequant_fp8_to_bf16.py

@@ -210,6 +210,11 @@ def dequantize_model(model_dir: str, out_dir: str):
             stats["scales_removed"] += 1
         
         out_path = os.path.join(out_dir, os.path.basename(f))
+        if os.path.exists(out_path) and os.path.getsize(out_path) > 0:
+            # Resume: skip already-dequantized shards
+            print(f"[{i+1}/{total_shards}] Skipping (already done): {os.path.basename(f)}")
+            del tensors, scales_in_shard
+            continue
         save_file(tensors, out_path)
         
         shard_time = time.time() - shard_start