Expand SFTTrainer API so Compute Can be Customized

**What did you find confusing? Please describe.**
How do I configure the number of instances trained with? What about the timeout? Is there a link to the implementation? Are there docs for the arguments it takes in for the constructor and train methods?

**Describe how documentation can be improved**
Add details about the arguments it takes in including examples

**Additional context**
There needs to be ways the above settings can be configured. For larger fine-tuning jobs it's too slow.

Start by locating the SFTTrainer constructor and train method implementation and their existing documentation. Document the arguments with examples, including the requested instance-count and timeout settings, and link to the implementation; the documentation should make these options and their configuration clear.

由索引模型根据 Issue 内容生成。