LLama3 MVP

[ngc92]

This implements the minimum necessary changes to get an implementation of LLama3 that is functional.
Key fixes compared to the current llama3 branch:

• ignore bias terms during backward
• ensure learning rate matches pytorch reference
• fix gradient checking (Fix gradient tests #801)

When trying to set up CI, we run into the problem that even the 1B model is too large to fit for training; I've tried two different things:
a) run reference code with --device=cpu. This works, but we see numerical differences quite prominently, would need to increase tolerances by ~10x for fp32 mode
b) use torchao's CPUOffloadOptimizer. Works, but introduces another dependency on the python side. Also changes the numerics, but only for the optimizer, so it doesn't break the gradient step. EDIT: CPUOffloadOptimizer is not compatible with gradient clipping, so I had to add a terrible hack :( The loss values set as targets in test_llama3.cu are generated from the .py file, but I ran it on a larger GPU so that these are without offloading.

What is still missing:

• cuDNN support
• tied weights (i.e., actually reproducing 1B/3B models)
• evaluation in the C training loop
• generating models from scratch
• haven't checked the non-HF code path for the python reference

Q: Do we really want to store the hidden dimension size as a floating-point factor , followed by 1024-roudning? Instead of just specifying hidden_dim explicitly? The factor would make sense if it was the same across all models, but it differs.

I do have code for all of these, but I'd like to keep the PRs at a managable size, so these changes aren't included here.