Dynamically generate FP8 configs for Fuji

[andersensam]

This PR aims to simplify the usage of FP8 for Fuji models. It removes the necessary edits previously required to enable FP8 support and now dynamically generates FP8 configs alongside single-host configs. This also extends FP8 support beyond GCP instance types.

Usage:

python3 -m axlearn.common.launch_trainer_main \
        --module=text.gpt.c4_trainer --config=fuji-7B-v2-flash-fp8 \
        --trainer_dir=<trainer_dir> \
        --data_dir=gs://axlearn-public/tensorflow_datasets \
        --jax_backend=gpu \
        --mesh_selector=gpu-a3-megagpu-8g-256 \
        --trace_at_steps=5

[andersensam]

Golden configs still need to be updated -- will run additional testing later and update accordingly

[hanzhi713]

LGTM

[markblee]

Will be great if we can limit the number of golden configs.