Add NEFTune implementation for Noised Embedding Instruction Fine-Tuning support

[arnavgarg1]

This PR adds an implementation of the original NEFTune paper to add noise to the embedding outputs at train time. Empirical evidence suggests that this improves model performance by up to 35% and reduces overfitting.

To use this in Ludwig with LLMs, just set neftune_noise_alpha in the model_parameters section of your LLM model config to a non-zero value.

model_type: llm
base_model: meta-llama/Llama-2-7b-hf
model_parameters:
    neftune_noise_alpha: 5

---

Implentation

I implemented this is as a forward hook that gets added to the embedding layer once we're inside the trainer (well after the model is already instantiated) and removed at the end of training. The base class abstraction allows us to modify the behavior in custom ways for this hook as well as future hooks, including using backward or forward hooks, or even not removing the hook at the end of training (if desired).

[github-actions]

Unit Test Results

  6 files  ±0    6 suites  ±0   14m 23s ⏱️ ±0s
12 tests ±0    9 ✔️ ±0    3 💤 ±0  0 ❌ ±0 
60 runs  ±0  42 ✔️ ±0  18 💤 ±0  0 ❌ ±0 

Results for commit b7e32a1. ± Comparison against base commit 2f159b6.

♻️ This comment has been updated with latest results.

[tgaddair]

Looks great! Just a couple nits.