include: remove config.h from headers

[rauteric]

Header files shouldn't include config.h

Issue #, if available:

Description of changes:

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.

[bwbarrett]

We're probably solving the wrong problem here, but this also isn't wrong, so approved.

[aws-nslick]

I'm somewhat against this. Some of these are just unnecessary, but some are not, eg: HAVE_DECL_FI_MR_DMABUF must be defined for nccl_ofi_mr.h to do the right thing.

I don't see any good reason why it should be possible to include nccl_ofi_mr.h without that being defined. We've been bit by this before.

The cleanest way to do this IMO is just to get rid of config.h entirely by removing AC_CONFIG_FILE(config.h) (or whatever) from configure.ac, so that -DHAVE_DECL_FI_MR_DMABUF=1 is passed on every invocation of the compiler and you never have to worry about config.h being the first include or making sure that every file is including it in the first place. I've proposed that before but it wasn't popular.

So as long as we have config.h, and it needs to be included by someone (either the TU itself, where every header relies on a transitive previous inclusion of config.h, or by the header itself pulling in config.h when it needs some definition from config.h), I think the latter is preferable and the file that uses it should be responsible for including it -- mostly based on arguments from Why Include What You Use? -- it's really nice to be able to grep for a header and see at a glance how files connect to each other.

There's an argument to be made that project public headers should be stateless and not have any configuration, but that's not the state of the project and removing the include from the header doesn't change that, it just means that there's a transitive* inclusion of config.h that must occur instead.

[rauteric]

Nick convinced me that this is not sufficient to deal with our benchmark code properly, so I will come up with something else that doesn't require this patch.

I think right now we're a bit inconsistent with whether we include config.h in the headers or not, but I'm fine leaving well enough alone.

[aws-nslick]

A portion of this commit was fine with me (just in terms of removing totally unnecessary #includes, ie: the one in the tuner).

If there's buy in from both of you, I'd like to just get iwyu running in github actions on every commit similar to other compile-time checks we have. There's a million ways to solve header consistency, but at least this would be automatic and consistent, and we wouldn't ever have to think about it again.

I'd actually proposed this previously in #351 (0e2d51d and f34a7ae)

[bwbarrett]

Unless we need a ton of workaround pragmas, I'm fine adding IWYU in the list of compile time checkers we run on every PR. I did try it with the version included in U20 without success, but didn't put a lot of time into it.