In our main branch, the C++26 reflection benchmarks were outdated (and wrong). This PR merges the changes @FranciscoThiesen's separate branch (francisco/ablation_study) with a few additional fixes:

• The file benchmark/static_reflect/citm_catalog_benchmark/rapidjson_citm_catalog_data.h did not build and appeared to be out of sync. Basically, some values have been optionals, and the code did not reflect that.
• Though it is not part of our public API, one overload to the simdjson::builder::to_json function returned a simdjson_error (which is an exception) instead of a simdjson::error_code (which is just a number, essentially). This was a bug in the main library (although unlikely to affect anyone at this time since it is not a documented function). Of course, I caught this because because I added checks for the return values (the error).
• I have updated our p2996/README.md so that it is clearer.
• I have also switched some simdjson functions from exception-based to exception-less. (This is not very important).

I note that Nlohmann parsing is disabled for the twitter dataset.

Note that @FranciscoThiesen's branch is not quite in sync as far as the main library is concerned, so there are some differences, possibly.