Parallel-For jobs

[LilithSilver]

Relies on #24 and predecessors, so let's resolve those first.

This PR implements IJobParallelFor with a basic work-stealing algorithm. The underlying data structure is like Chase-Lev's work-stealing deque, except I made a special version that just operates on a Range of ints, giving both O(1) time and space complexity.

The work-stealing algorithm itself is nothing special; thieves just try until all the deques are empty and then give up. There's no yielding or sleeping because we know no parallel ranges can spawn other parallel ranges, so we know we want to use the full system resources.

Here's a simple increment-counter benchmark:

Method	Threads	MaxConcurrentJobs	Mean	Error	StdDev	Allocated
BenchmarkForLoop	0	32	68.05 ms	0.191 ms	0.149 ms	600 B
BenchmarkStandardParallelFor	0	32	446.61 ms	0.249 ms	0.233 ms	150080 B
BenchmarkParallelFor	0	32	467.79 ms	6.362 ms	5.640 ms	600 B

And here's testing multiplying a 1024x1024 matrix (an actually difficult task):

Method	Threads	MaxConcurrentJobs	Mean	Error	StdDev	Median	Allocated
BenchmarkForLoop	0	32	273.91 ms	0.750 ms	0.701 ms	274.07 ms	600 B
BenchmarkStandardParallelFor	0	32	34.35 ms	0.167 ms	0.148 ms	34.35 ms	5496 B
BenchmarkParallelFor	0	32	47.16 ms	2.196 ms	6.474 ms	51.10 ms	600 B

As you can see, it hits in a similar ballpark to Parallel.For. It's a bit slower, but I think it's "good enough" and I'm not sure how to improve it further.