[float8] Support power of 2 scales with PerRow scales for inference

[danielvegamyhre]

Summary

• float8 training w/ rowwise scales uses power of 2 scales by default, to reduce quantization error
• float8 inference w/ Float8DynamicActivationFloat8WeightConfig using PerRow scaling doesn't support power of 2 scales
• users have reported they want to be able to use power of 2 scales for inference after training with them.

cc @drisspg @vkuzo