Restructure multithreading

[HeroicKatora]

Redefine the Worker trait, which allows the chosen worker to create a
new scope. This is relevant for a newly created (and installed) rayon
worker.

The rayon thread pool is now local to the decoding step. This fixes an
issue where improper task scheduling would deadlock decoding. It's not
clear how the intended task scheduling can be reliably achieved without
the guarantee of having at least a second, free, worker thread.

Closes: #227

[HeroicKatora]

Added more tests, no based off the global pool, to demonstrate/test that this does not make a quantifiable difference. As suggested to add them here. @awused Did I miss any, do you have more suggestions?

[HeroicKatora]

Okay, scratch that. This is why it was a draft. How about this: In this rayon case, we always in_place_scope the main decoder loop. (The architecture should already be there after the rework). Then, instead of creating separate, parallel worker tasks, in append_row we run the Immediate worker as a new scoped task—this is already guaranteed non-blocking, still single threaded, but I'll get to that. Both start and get_result will run directly on the immediate worker without any of dubious communication structure of the multithreaded module.

Now the critical part: We modify Worker::append_row such that multiple rows can be spawned at the same time. The rayon implementation can then create multiple tasks in its scope, which lets other threads worksteal from us. Not sure about the runtime implications of this scheme. However, this way making the rayon worker entirely non-blocking might be easier than imagined.

[HeroicKatora]

I've given rayon usage a complete overhaul. No more blocking at all but Mutex to modify the shared result sets.

[HeroicKatora]

Note: this is not final, but setup to always choose rayon over multithreading over immediate. Note that we may choose any available one at runtime with this style.