Collision detection optimizations, benchmark runner, and collision detection benchmark

[cspotcode]

This is related to #1504. Where the other PR is a draft which experimented with many ideas, this PR includes only the changes that had the biggest positive improvements. It also adds the benchmark runner and collision benchmark.

I think this PR is suitable to be merged as-is.

Here are my benchmark results. I ran it against each commit twice, I'm not sure if that was necessary but I was worried that my laptop would always run the first commit fastest because it was not hot (overheating) yet. That may have been wrong or unnecessary.

Command	Mean [s]	Min [s]	Max [s]	Relative
0066a6ae _check_for_collision: optimize broadphase	30.941 ± 0.864	30.330	31.552	1.01 ± 0.03
6c023813 get_adjusted_hit_box: list comprehension into tuple is faster than generator into tuple	36.530 ± 0.386	36.257	36.803	1.19 ± 0.01
5748dec0 get_adjusted_hit_box: inline rotate_point() call, avoid attribute lookups on self and math	36.603 ± 0.262	36.418	36.789	1.19 ± 0.01
0066a6ae _check_for_collision: optimize broadphase	30.664 ± 0.028	30.644	30.683	1.00
6c023813 get_adjusted_hit_box: list comprehension into tuple is faster than generator into tuple	36.045 ± 0.347	35.800	36.291	1.18 ± 0.01
5748dec0 get_adjusted_hit_box: inline rotate_point() call, avoid attribute lookups on self and math	40.900 ± 5.103	37.291	44.508	1.33 ± 0.17

[pvcraven]

I think the speed-ups would be good. I'm inclined to merge those in.

I'm not sure about leaving the benchmarks directory in.

• Either leave it in and add a readme documenting how it works, using it, sample output
• Or just drop it if documenting it is too burdensome.

[einarf]

As discussed on discord we merge this as is. The contents in the benchmarking directory we can change to using standard cProfile using the build in profiler in arcade. Optionally just use simple timeit. Instead of bringing in large examples we can test individual functions only.