[docs] training performance

[stevhliu]

• created a new performance section divided into memory and speed optimizations
• model memory training anatomy guide is now the more descriptive and simplified GPU memory usage doc
• add brief loss scaling section in gradient accumulation for custom loss function users

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.