- Allow quickly switching between version tags

#3150 (comment)

To support [flights-200k.arrow](https://github.com/vega/vega-datasets/blob/f637f85f6a16f4b551b9e2eb669599cc21d77e69/data/flights-200k.arrow)

Not required for these requests, but may be helpful to avoid limits

As an example, for comparing against the most recent I've added the 5 most recent

See https://docs.github.com/en/rest/repos/repos?apiVersion=2022-11-28#list-repository-tags

- Basic mechanism for discovering new versions
- Tries to minimise number of and total size of requests

Experimenting with querying the url cache w/ expressions

- `metadata_full.parquet` stores **all known** file metadata
- `GitHub.refresh()` to maintain integrity in a safe manner
- Roughly 3000 rows
- Single release: **9kb** vs 46 releases: **21kb**

https://github.com/vega/altair/actions/runs/11495437283/job/31994955413

- Still undecided exactly how this functionality should work
- Need to resolve `npm` tags != `gh` tags issue as well

Suffix was only added due to *now-removed* test files

Will be following up with some more splitting into composite modules

Handled in `Application` now

`tools.datasets.npm` will later be performing the requests that are in `Dataset.__call__` currently

-`data.name()` -> `data(name)`
- `data.name.url` -> `data.url(name)`

Since we're just using strings, there is no need for 2 forms of the name.
The legacy package needed this for `__getattr__` access with valid identifiers

I think this was added in error, since the schema of the file never had `semver` columns

Only noticed the bug when doing a full rebuild

Temporary home, see module docstring

- Much more generic now in what it can be used for
- For the caching, I'll need more columns than just `"url_npm"`
- `"url_github" contains a hash

- Common prefix to all rows in `metadata[url_github]`
- Stripping this leaves only `sha`
- For **2800** rows, there are only **109** unique hashes, so these can be used to reduce cache size

Related 661a385

Not planning to support this anymore, requires `fsspec` which isn't in `dev`

```
InvalidInputException
Traceback (most recent call last)
Cell In[6], line 5
       3 with duck._reader._opener.open(url) as f:
       4     fn = duck._reader._read_fn['.json']
----> 5     thing = fn(f.read())

InvalidInputException: Invalid Input Error: This operation could not be completed because required module 'fsspec' is not installed"
```

Will put this in another PR, but need it here for IDE support

- `tools.datasets`: Building & updating metadata file(s), generating annotations
- `altair.datasets`: Consuming metadata, remote & cached dataset management

Since `metadata.parquet` is sorted, this was already the behavior when not providing a tag

Still on the fence about `Loader.cache_dir` vs `Loader.cache`

Also incorporates changes from previous commit into `__repr__`
4a2a2e0

Initially quite basic, need to add more parameterize and test caching

Possibly fix https://github.com/vega/altair/actions/runs/11768349827/job/32778071725?pr=3631

Forgot that this gets uninstalled in CI
https://github.com/vega/altair/actions/runs/11768424121/job/32778234026?pr=3631

- Rename and properly document "file-like object" handling
  - Also made a bit clearer what is being called and when
- Use a more granular approach to skipping in `@backends`
  - Previously, everything was skipped regardless of whether it required `pyarrow`
  - Now, `polars`, `pandas` **always** run - with `pandas` expected to fail
- I had to clean up `skip_requires_pyarrow` to make it compatible with `pytest.param`
  - It has a runtime check for if `MarkDecorator`, instead of just a callable

bb7bc17, ebc1bfa, fe0ae88,
7089f2a

- Moved `_Reader._metadata` -> module-level constant `_METADATA`.
  - It was never modified and is based on the relative directory of this module
- Generally improved the readability with more method-chaining (less assignment)
- Renamed, improved doc `_filter_reduce` -> `_parse_predicates_constraints`

Still need examples for `Loader.__call__`

`get_` was the wrong term since it isn't a free operation

There are enough alternatives here, it only added complexity

Disabled by default, since there are 74 datasets