Purge INT4 references — expert weights are FP4 (E2M1), not INT4

All docs and scripts updated. Historical memory entries annotated.

scripts/model_opt_nvfp4_full.py

@@ -4,7 +4,7 @@ ModelOpt NVFP4 quantization — full model.
 
 Quantizes ALL weights (attention + experts + shared MLP) to NVFP4.
 Requires a pure BF16 source model (from scripts/dequant_fp8_to_bf16.py)
-to avoid FP8/INT4 kernel issues on Blackwell GPUs.
+to avoid FP8/FP4 kernel issues on Blackwell GPUs.
 
 Available NVFP4 quantization strategies (from modelopt huggingface_example.sh):
   - nvfp4               : Full model NVFP4 quantization (this script)