Backend fastpath

[slevang]

• Closes First invocation of open_dataset takes 3 seconds due to backend entrypoint discovery being slow #10178
• Tests added
• User visible changes (including notable bug fixes) are documented in whats-new.rst

Implements this proposal, bypassing the list_engines call in the case that an explicit string engine is passed that exists in the standard list. This can shave up to a few seconds off the first open_dataset call depending on your env size.

Only side effect I can see is it does change the error reporting for missing engines. Now we get:

ImportError: The zarr package is required for working with Zarr stores but could not be imported. Please install it with your package manager (e.g. conda or pip).

Instead of:

ValueError: unrecognized engine 'zarr' must be one of your download engines: ['netcdf4', 'h5netcdf', 'scipy', 'cfgrib', 'gini', 'rasterio', 'store']. To install additional dependencies, see:
https://docs.xarray.dev/en/stable/user-guide/io.html 
https://docs.xarray.dev/en/stable/getting-started-guide/installing.html

This also doesn't help for non-standard engines, e.g. open_dataset(..., engine="cfgrib"), but if speed is crucial you can pass the backend object itself in that case.

[jsignell]

The new error message kind of seems better than the old one honestly for engine='zarr'. Do you think it's worth adding a test for this change? Would it be too heavy to try to have a test that proves that for these default engines you never need to call list_engines?

[jsignell]

This looks great! I love the test 😍