Simplify tag_invoke and friends

[hkaiser]

The goal is to streamline the implementation of tag_invoke and friends.

@msimberg, @K-ballo, if you could spare a minute, I'd appreciate a second pair of eyes as I'm not sure that the proposed changes will not subtly change overload resolution.

[hkaiser]

@msimberg not sure where the 443_executor test problem comes from, but otherwise things seem to be fine now.

[hkaiser]

retest lsu

[msimberg]

@msimberg not sure where the 443_executor test problem comes from, but otherwise things seem to be fine now.

Not sure since it's probably from the previous CI run, but most likely it's the same failure that was mentioned earlier and is fixed by #5274?

[msimberg]

@hkaiser would you mind summarizing the changes here. It looks like the tag_X_invoke_ns namespace became regular (non-inline) namespaces, the helper base classes moved into sub-namespaces, while being typedefed in separate inline namespaces. Is this to prevent certain accidental/unwanted overloads matching or similar? Is there some other change that I've missed?

[hkaiser]

@hkaiser would you mind summarizing the changes here. It looks like the tag_X_invoke_ns namespace became regular (non-inline) namespaces, the helper base classes moved into sub-namespaces, while being typedefed in separate inline namespaces. Is this to prevent certain accidental/unwanted overloads matching or similar? Is there some other change that I've missed?

Yes, this mostly sums it up. The main change is that the helper base classes now directly dispatch to the tag_invoke implementations instead of using hpx::functional::tag_invoke as before. This saves one function call for each tag_invoke invocation. The namespace changes were needed to avoid the roundtrip through the hpx::functional::tag_invoke facility.

[msimberg]

Thanks @hkaiser! I think we can ignore the ICC failures. As far as I can tell they only happen because this is still using the configuration from before #5199 was merged.