[doc] Document the data machinery

[jlamypoirier]

We need to document the new dataa mechanism, introduced in #37, #40, #104, etc.

The doc should at least describe:

• The different types of datasets (samplable, sampled, indexed, etc.)
• The different dataset classes available for GPT, and generic ones.
• The dynamic dataset class instantiation mechanism (to move elsewhere if/once we generalize)
• How datasets are put together in Data
• Maybe also the dataset preparation mechanism.

It should let the user know how to:

• Prepare a dataset with the prepare command (recipe is already in the doc, might just need to point to it)
• Configure Fast-LLM for typical dataset use cases (ex. dataset in multiple files, train/valid/test split, blending multiple datasets, etc.)
• Extend GPT with a custom configurable dataset class
• Extend GPT data and/or datasets for a custom model (ex. gpt_custom)
• (Advanced, can be postponed) Create a data and dataset for a custom model that doesn't extend GPT, reusing the existing generic machinery. (Need to generalize dynamic configuration first)
• (Can be postponed) Create a custom data that doesn't reuse the dataset machinery. Easy but not sure how much we need it.

[tscholak]

Hey @jlamypoirier,

Let's move fast on this. The immediate need is a set of practical examples that show how different dataset configurations achieve useful, real-world outcomes. No deep dives, no background, no code, just clear, runnable configs with comments explaining what each one does and why someone might use it.

This should be achievable within a day.

Please sync with @RaymondLi0 before starting to identify 4 to 5 meaningful, distinct examples. The goal is for someone new to Fast-LLM to copy a config, tweak a few parameters, and get results without needing to read a wall of text.

We'll worry about more detailed explanations later. Right now, examples > reference handbook.

Thanks!