A few changes to ``config.set``

[jcrist]

The main intent of this PR is to support setting config values through
keyword parameters instead of requiring a dictionary for nested config
values. We follow the same convention as we do for environment
variables, where double-underscores are interpreted as a ., and
indicate an increase in nesting level. Now the following are equivalent:

with dask.config.set({"foo.bar.baz": 1}):
    ...

with dask.config.set(foo__bar__baz=1):
    ...

If a mapping argument and keyword arguments are both provided, the
mapping is applied first, then the keyword arguments (the keywords are
effectively overrides for the mapping).

In the process, I discovered a bug in the rollback code when dealing
with multiple keys interacting with the same nested value. To deal with
this I rewrote the rollback code in a way that was simpler for me to
understand, and hopefully others as well.

• Tests added / passed
• Passes black dask / flake8 dask

[jcrist]

cc @djhoese, this may interest you.

[jcrist]

Failure is unrelated, and may be fixed by #5228.

[djhoese]

Thanks for the mention. This makes sense to me from a higher level. I'm not super excited about having to mirror the rollback changes to donfig, but that's my fault for not finishing my PRs getting dask on an object oriented config object.

[djhoese]

You mentioned a bug in the rollback, does your new test cover this case?

[jcrist]

Yeah, the issue had to do with multiple keys in a set operation interfering, mixing keywords and the mapping handles this case. FWIW I don't think this is a bug that could easily be hit without allowing mixing mapping and keyword arguments (and still is unlikely to be hit). I find the new rollback code easier to understand, and it's deterministic (we iterate through a list in the reverse order of operations, not through a dict which may be unordered depending on python version).

[jcrist]

@mrocklin, quick review? Planning to merge this later today otherwise.

[mrocklin]

Can I trade you for reviewing dask/distributed#2913 ?

[mrocklin]

The support of __ looks good to me. The rewrite of writing and unrolling config I understand less, but it all seems reasonable. I'm +1

[mrocklin]

The support of __ looks good to me. The rewrite of writing and unrolling config I understand less, but it all seems reasonable. I'm +1