Support Granite 3.0 and 3.1 models

[JamesKunstle]

Granite 3.(0,1) models are Llama-architecture models with some different scaling terms in various places. This commit adds granite model patching for decoder-only granite 3 models (not multimodal) and the corresponding tests.

Summary

This change enables patching Granite 3.(0,1) models w/ Liger kernels. We would like to use Liger kernels in our training implementation but we're a Granite-first codebase for the moment.

Testing Done

Convergence tests confirm that loss and model parameters are equivalent w/ and w/o Liger kernels. Logits, however, are not equivalent even when only swapping the SwiGLUMLP layer. The ator and rtol may need to be tuned for Granite vs. Llama, I'm going to continue investigating this before this PR is merged.

• Hardware Type: EC2 g6e.12xlarge; 4xL40s
• run make test to ensure correctness
• run make checkstyle to ensure code style
• run make test-convergence to ensure convergence

[JamesKunstle]

Fixes #557