Caching cudaFunctorAttributes and whether L1/Shmem prefer was set

[crtrott]

This reduces launch latencies, and closes the gap between a raw CUDA
launch and Kokkos Launch to about 0.2us (empty Kernel parallel_for).

Test case comparison the different times are non-fenced, and fenced after each kernel call. In parenthesis fenced with an extra fence after the Repeats loop (doesn't matter, since I can't enqueue so many kernels anyway, so it adds like 5us/Repeats).

Old:

LoopCount FunctorCArraySize InnerLoopCount Repeats
10000 1 1 10000 raw: 3.360421 (3.360927) parallel_for: 4.337152 9.894000 ( 4.337722 9.894113 ) parallel_reduce: 14.843851 15.885946 ( 14.844018 15.886035 ) parallel_reduce(view): 5.188312 12.267948 ( 5.779884 12.268049 )
10000 16 10 10000 raw: 3.291513 (3.292117) parallel_for: 4.385937 9.595486 ( 4.386451 9.595609 ) parallel_reduce: 14.408576 15.886102 ( 14.408739 15.886200 ) parallel_reduce(view): 5.352994 12.819591 ( 5.967374 12.819692 )
10000 200 10 10000 raw: 3.517541 (3.518094) parallel_for: 10.905219 12.708534 ( 10.907358 12.708635 ) parallel_reduce: 18.430118 19.563913 ( 18.430311 19.564008 ) parallel_reduce(view): 14.289269 15.626251 ( 14.290132 15.626348 )
10000 3000 10 10000 raw: 3.544100 (3.544642) parallel_for: 13.466383 15.665507 ( 13.466974 15.665618 ) parallel_reduce: 22.444786 23.621668 ( 22.444963 23.621766 ) parallel_reduce(view): 16.727474 20.115662 ( 16.728418 20.115771 )

New

LoopCount FunctorCArraySize InnerLoopCount Repeats
10000 1 1 10000 raw: 3.388423 (3.388938) parallel_for: 3.611583 9.004801 ( 3.612182 9.004904 ) parallel_reduce: 13.554614 15.010310 ( 13.554782 15.010404 ) parallel_reduce(view): 5.208067 11.574165 ( 5.807058 11.574267 )
10000 16 10 10000 raw: 3.320113 (3.320646) parallel_for: 3.703650 8.824116 ( 3.704197 8.824227 ) parallel_reduce: 13.593689 15.010715 ( 13.593853 15.010810 ) parallel_reduce(view): 5.392943 12.073576 ( 6.007791 12.073711 )
10000 200 10 10000 raw: 3.537305 (3.537863) parallel_for: 10.890026 12.105710 ( 10.890533 12.105827 ) parallel_reduce: 17.602991 18.618286 ( 17.603149 18.618372 ) parallel_reduce(view): 14.308791 14.845797 ( 14.309610 14.845891 )
10000 3000 10 10000 raw: 3.538977 (3.539568) parallel_for: 13.337157 14.805056 ( 13.337811 14.805167 ) parallel_reduce: 21.849866 22.893569 ( 21.850049 22.893656 ) parallel_reduce(view): 16.641581 19.349426 ( 16.642512 19.349533 )