feat(cuda): make GPU dispatch mode configurable

[0ax1]

Allow for explicitly choosing dyn dispatch or standalone kernels. We primarily do this to be able to test and bench the standalone CUDA kernels.

In a follow up, I'll change the logic to prefer standalone kernels for auto mode, if available for an array tree, over dyn dispatch.

[0ax1]

Proof by bench:

Warning: It is not recommended to reduce nresamples below 1000.
Benchmarking bitunpack_cuda_u8/bitunpack/3bw: Warming up for 1.0000 ns
Warning: Unable to complete 10 samples in 1.0ns. You may wish to increase target time to 4.2s.
bitunpack_cuda_u8/bitunpack/3bw
                        time:   [72.466 µs 78.387 µs 78.355 µs]
                        thrpt:  [1188.6 GiB/s 1188.1 GiB/s 1285.2 GiB/s]
                 change:
                        time:   [-51.700% -44.557% -44.078%] (p = 0.00 < 0.05)
                        thrpt:  [+78.822% +80.364% +107.04%]
                        Performance has improved.
Found 1 outliers among 10 measurements (10.00%)
  1 (10.00%) high severe

Benchmarking bitunpack_cuda_u16/bitunpack/5bw: Warming up for 1.0000 ns
Warning: Unable to complete 10 samples in 1.0ns. You may wish to increase target time to 4.2s.
bitunpack_cuda_u16/bitunpack/5bw
                        time:   [82.442 µs 82.816 µs 83.530 µs]
                        thrpt:  [2229.9 GiB/s 2249.1 GiB/s 2259.3 GiB/s]
                 change:
                        time:   [-39.171% -38.913% -38.575%] (p = 0.00 < 0.05)
                        thrpt:  [+62.801% +63.702% +64.394%]
                        Performance has improved.