Update README.md for mx_formats build from source (#2934)

Update README.md

Author: vkuzo

torchao/prototype/mx_formats/README.md

@@ -36,6 +36,8 @@ including [downloading a tokenizer](https://github.com/pytorch/torchtitan?tab=re
    - mxfp8_cublas: `TORCHTITAN_ROOT=<path> MX_RECIPE="mxfp8_cublas" ./benchmarks/float8/training/llama3.sh`
    - mxfp8_cublas_rceil: `TORCHTITAN_ROOT=<path> MX_RECIPE="mxfp8_cublas_rceil" ./benchmarks/float8/training/llama3.sh`
 
+> :warning: For now you need to build `torchao` from source for optimal training performance. See https://github.com/pytorch/ao/issues/2932 for details.
+
 # User API
 
 ## MX training