Currently, `dataset.py` is 10956 lines long. This makes doing any work with current LLMs basically impossible — with Claude's tokenizer, the file is 104K tokens, or >2.5x the size of the _per-minute_ rate limit for basic accounts. Most of xarray touches it in some way, so you generally want to give it the file for context.

Even if you don't think "LLMs are the future, let's code with vibes!", the file is still really long; can be difficult to navigate (though OTOH it can be easy to just grep, to be fair...).

So I would propose:
- We start breaking it up, while also being cognizant that big changes can cause merge conflicts
- Start with the low-hanging fruit
  - For example, this PR moves code outside the class (but that's quite limited)
  - Then move some of the code from the big methods into functions in other files, like `curve_fit`
- Possibly (has tradeoffs; needs discussion) build some mixins so we can split up the class, if we want to have much smaller files
- We can also think about other files: `dataarray.py` is 7.5K lines. The tests are also huge (`test_dataset` is 7.5K lines), but unlike with the library code, we can copy out & in chunks of tests when developing.

(Note that I don't have any strong views on exactly what code should go in which file; I made a quick guess — very open to any suggestions; also easy to change later, particularly since this code doesn't change much so is less likely to cause conflicts)