Performance overhaul for table.update()

[texodus]

This PR contains a number of fixes applied as a result of some runtime profiling of the table.update() mechanism:

• Removed std::unordered_map in favor of tsl::hopscotch_map, which improves primary key lookup performance, and overall update processing time is improved by 50%.
• Added throttling to this method. This means that on_update() is no longer called 1:1 with calls to update() (though the final on_update() will always be consistent with table state). As a result, update() calls will no longer stack up and overwhelm other API calls, which leads to drastic improvement in perceived performance, and much smarter interleaving of update() and other calls which reflect the view() to the <perspective-viewer> element.

Additionally, a few other enhancements have been made which facilitated these principal fixes:

• Benchmarks now mark outlier observations (more accurate benchmarks).
• CMake executable builds are now parallelized (faster builds).
• yarn test now respects -t in focus mode (faster test iteration).