slim_model: options for float16, checkpoint selection

[nostalgebraist]

Contains some additions I made to slim_model.py when I was using it with a finetuned model.

• --f16: converts to float16

  • Why: Since many people are doing inference with float16 weights, it seems better to just go directly to float16, without the bfloat16 step

• --ckpt-step: passing e.g. --ckpt-step 401 will load from step_401/ even if it's not the most recent checkpoint

  • Why: I wanted to convert the model with the lowest val loss, but usually there were 1 or more checkpoints after that step

[kingoflolz]

Otherwise looks good!

[kingoflolz]

Thanks!