Support Constant Learning Rate with Cooldown

[LoserCheems]

Feature request

In transformers.optimization support constant learning rate with cooldown functions.

Motivation

This method will implement that scaling experiments can be performed with significantly reduced compute and GPU hours by utilizing fewer but reusable training runs.
SmolLM had used this method to train a series of SOTA small language models.
Paper: https://arxiv.org/pdf/2405.18392

Your contribution

I've created a branch, I'm finishing the implementation of these functions, and intend to submit a PR.