Add bitnet

[dhiaEddineRhaiem]

This PR adds support for BitNet Pretraining for gpt-like architectures:

1. Integrates BitNet layer specs into GPT model building pipeline via --use-bitnet flag (requires --transformer-impl local)
2. Implements BitNetColumnParallelLinear and BitNetRowParallelLinear layers that apply 1-bit weight quantization
3. Uses onebitllms Triton kernels for efficient activation and weight quantization during forward pass

cc @sbhavani

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[janEbert]

Hey @dhiaEddineRhaiem, sorry for not getting to this earlier. I'm a bit unsure about the dependency you're introducing. It has already not been updated for 9 months, so its future support seems very unstable.

[janEbert]

I'll discuss the introduction of the dependency with our team. We may take some time to properly review, but we'll try to up the pace. :)

[Phlip79]

We are changing our review process and marking all open, unlabeled PRs as draft. This change will go in effect starting once #3659 is merged.

Moving forward, all PRs will be required to start as draft PRs. If you wish to get your PR merged, mark your PR as “Ready for review”. Read more about the new process at submit.md.