Overhaul logging mechanism

[psalz]

This changes the logging mechanism used throughout the runtime from having to pass around the detail::logger class to simple free-standing macros such as CELERITY_INFO(...). While the new mechanism no longer allows for arbitrarily nested "contexts", a single context can be set through CELERITY_SET_LOG_CTX (currently only used in worker jobs).

Full list of changes:

• Get rid of JSON output for trace log level
• Remove logger class, use macros instead
• Only a single level of context supported
• Adjust log level for many messages, distinguish between debug/trace
• Add additional information in various logging messages
• Default log level to info (release builds) and debug (debug builds)

[fknorr]

Thank you, human readable logs are a very welcome change! A few things:

[fknorr]

Thanks! A few more things, mostly about includes. I would prefer if we kept iostreams out of as many translation units as possible because it tends to be quite expensive to compile.

[psalz]

I've addressed the additional feedback by @fknorr and in that process also made some improvements to the user_bench utilities, along with some breaking changes (the whole thing is experimental and we don't use it in any of the examples currently, so I think it should be fine).

As mentioned in the comment above, I've mainly gotten rid of the log_map type as an alias for std::unordered_map, both to address limitations in the types that can be used, as well as to avoid the unnecessary overhead of creating a whole hash map only for it to be immediately turned into a string. Note that the log_map type still exists, but only as a helper to specialize fmt::formatter. While I could've moved the formatting of the std::tuple directly into the constructor of log_context, I didn't want to repeatedly call fmt::format and then concatenate everything into a stringstream because that seemed wasteful.

[fknorr]

Thank you! After some digging, the tie-in between fmt and iostreams in our code, aka <fmt/ostr.h> has two origins:

• print_utils.h and .cpp expose ostream& operator<< for the subrange and chunk types, which are internally implemented using fmt, so we are going back and forth between fmt and stringstreams.
• The Allscale headers define ostream& operator<< on the GridBox type. Its only use is in region_map_testspy::print_regions, and that is currently dead code as the only (transitive) invocation is commented out in graph_compaction_tests:142.

However, the major offender now is the Allscale assertions header which pulls in <iostreams> unconditionally. I feel like this is beyond the scope of this PR now, so I'll cease my iostreams crusade.