Run duck array tests in separate CI jobs

[dcherian]

What is your issue?

As a follow up to #9798 we should consider running a two CI jobs with pint, sparse, jax installed (one with latest version, one with minimum supported version), and remove them from the other jobs.

cc @slevang

[slevang]

It looks like sparse and pint are installed "as able" in the various CI environments (e.g. all-but-dask/numba, min-all-deps).

sparse is listed as an optional dependency in pyproject.toml, but is only imported twice in core code and only as a means of special casing some duck array handling thing, which could probably be removed. pint is not listed as an optional dependency.

My thought would be to:

1. remove pint/sparse/jax from the core CI environments
2. consider creating a new optional dependency group that gives some user guidance on the array frameworks we test and know mostly work. Today that would be array = ["sparse", "pint", "cupy", "jax"].
3. as suggested, create latest and minimum CI envs that test duck array integration stuff. Unless that should live externally (xarray-array-testing).

[dcherian]

Sounds OK to me.

I'd also suggest restricting the running tests to only those that actually test duck array integration to save some CI time.