Update_pretrain_benchmarks

[crinard]

Updates the benchmarking script for updates to LLM Foundry, Torch 2.0, Torch.compile, greater FSDP sweeps, allow H100 profiling and FP8, and add tables for H100 w/ BF16 & (very sparse) FP8.

[vchiley]

Comments about the table

Can we

• separate out the table into 2 tables. one for amp_fp16 and another for amp_fp8
• include a Model TFLOP column
  • MFU between different hardware is hard to compare
• amp_fp8 is only used with MP Mode: DEFAULT; why not use PURE?
• remove 350M results