[Meta] Improve contributor guidelines

Our current [contributing guidelines](https://metal.juliagpu.org/stable/faq/contributing/) have not been updated in a while, and while they provide good information for improving the low-level Julia -> Metal interface, there isn't really any information on what's expected/helpful when submitting and issue/PR.

I suggest adding this, as well as LLM guidelines, and moving that from the documentation into the repo so that it's easier to find.

Until I take the time to actually write them, here are some bullet points of things I think should be included. The guiding principle for all of these are to reduce back-and-forth between contributor and reviewer, and to speed up the process of getting a PR merged:

- LLM assistance policy: See below
- Formatting. These would also be added to the bot comment:
- **No** formatting of untouched existing code
- No formatting of modified code until it has been reviewed. (Ex. spacing between expressions. This makes it hard to differentiate between code changes and aesthetic changes)
- Recommend not using Runic for formatting the `@objc` calls.
- MTL, MPS, MPSGraphs wrappers: Add some guidance around the general organization/structure of the low-level wrappers/high-level wrappers.
- Staying mindful of our limited CI resources (only 5 apple-silicon runners in buildkite shared between the julia community)

### LLM use

Ghostty's [policy](https://github.com/ghostty-org/ghostty/blob/main/CONTRIBUTING.md#ai-assistance-notice) aligns with how I feel about it (minus the part about personally using them a lot).

The main points of the policy would be:
- Disclose any use of AI assistance and the extent to which it was used (with an exception for short tab-autocomplete snippets)
- The submitter is still expected to fully understand their submission as if they had written it themselves.
- This includes having gone over the Pull Request/Issue description for hallucinations/inconsistencies.
- It's better to say you don't understand something than to include LLM hallucinations. This makes it easier to figure out what's wrong in issues, and helps reviewer to guide the contributor.