bench: CUDA host-to-device copy modes

[0ax1]

Compare pageable host memory with cuMemHostAlloc pinned allocations using default flags and WRITECOMBINED.

Benchmark results on a GH200:

cuda/load_to_device/memcpy_htod/pageable/1GiB
                        time:   [10.717 ms 10.754 ms 10.793 ms]
                        thrpt:  [92.649 GiB/s 92.989 GiB/s 93.306 GiB/s]

cuda/load_to_device/memcpy_htod/pinned_default/1GiB
                        time:   [10.085 ms 10.265 ms 10.527 ms]
                        thrpt:  [94.992 GiB/s 97.423 GiB/s 99.159 GiB/s]

cuda/load_to_device/memcpy_htod/pinned_write_combined/1GiB
                        time:   [21.043 ms 21.127 ms 21.204 ms]
                        thrpt:  [47.161 GiB/s 47.333 GiB/s 47.522 GiB/s]

cuda/load_to_device/device_alloc_memcpy_htod/pageable/1GiB
                        time:   [42.625 ms 42.704 ms 42.781 ms]
                        thrpt:  [23.375 GiB/s 23.417 GiB/s 23.460 GiB/s]

cuda/load_to_device/device_alloc_memcpy_htod/pinned_default/1GiB
                        time:   [41.864 ms 42.186 ms 42.592 ms]
                        thrpt:  [23.478 GiB/s 23.704 GiB/s 23.887 GiB/s]
                 change:
                        time:   [+1.7580% +2.5859% +3.6570%] (p = 0.00 < 0.05)
                        thrpt:  [-3.5280% -2.5207% -1.7276%]

cuda/load_to_device/device_alloc_memcpy_htod/pinned_write_combined/1GiB
                        time:   [51.986 ms 52.077 ms 52.166 ms]
                        thrpt:  [19.170 GiB/s 19.202 GiB/s 19.236 GiB/s]

The insights here being, WRITECOMBINED yields significantly slower copy performance, whilst pageable host memory is roughly on par with pinned host memory (without using WRITECOMBINED).