Add MFU (Model FLOPs Utilization) logging

[xingyaoww]

This PR implements MFU (Model FLOPs Utilization) logging as requested in pytorch#2100.

Changes

1. Added MFU utility module (torchtune/utils/mfu.py) with functions to:

   • Calculate theoretical peak FLOPS for GPU
   • Calculate actual MFU percentage
   • Calculate model FLOPs for one forward pass

2. Modified LoRAFinetuneRecipeDistributed to:

   • Calculate model FLOPs after initialization
   • Log MFU alongside other metrics in training loop

The MFU metric is now logged with the same frequency as other metrics (controlled by log_every_n_steps) and is available in all supported logging backends (terminal, disk, WandB, TensorBoard).

Closes pytorch#2100