Improve performance of select-top-k WARP_SORT implementation

[achirkin]

A few simplifications and tricks to improve the performance of the kernel:

• Promote some constants to static constexpr
• Allow capacity < WarpSize
• Reduce the frequency of sort operations for filtered version
• Remove warp_sort::load to simplify the api and implementation

[achirkin]

NB from previous discussions: I considered merging done() and store(..) together, but decided against that for two reasons:

1. Atm, store(...) is rather fast and can be used to store the content multiple times, after done() has been called once. Looks like a useful behavior.
2. I couldn't come up with a clean solution without the use of virtual functions, but I'm afraid introducing one would incur extra overheads.

[achirkin]

NB: more tests are coming in #618

[tfeher]

Hi Artem, thanks for the PR! It looks good, just a found a few small things.

Question, is the following statement true? With these modifications, capacity < warpsize is allowed. Since the capacity is chosen automatically according to the k value, the existing test cover these changes.

Please add to the PR description that the warp_sort::load functions are removed to simplify the API.

[achirkin]

Question, is the following statement true? With these modifications, capacity < warpsize is allowed. Since the capacity is chosen automatically according to the k value, the existing test cover these changes.

Yes, that is true. Also a few more tests are coming in #618 (which expands the api a little by allowing to pass nullptr as input indices).

[tfeher]

Thanks Artem for addressing the issues! The PR looks good to me.

[cjnolet]

@gpucibot merge