Implement parallel logical algorithms.

[trivialfis]

Extracted from #8822 .

[RAMitchell]

Is this faster at all?

[trivialfis]

@RAMitchell Good point.

For simple conditions like v > 0.0f, the difference is negligible. For more complicated checks like whether the input is an integer and larger than 0 (NDCG label), the difference is somewhat visible but still not an issue.

std::vector<float> data(100000000);

parallel none_of:      0.059212s
std::none_of:  0.140824s

I simply implemented it as these two functions when used for data validation, it has to walk through the entire dataset. Seems not necessary.

[RAMitchell]

Feel free to merge if you want. My experience in the past is that simple operations on data don't benefit much. Multithreading then adds a bit of complexity and makes debugging harder.

[trivialfis]

Make sense, thank you for the review. Will remove the code from the LTR branch.