Performance Regression in v0.9.0

[d3v-null]

Recent changes to allow hyperbeam to run on AMD GPUs have impacted performance on CUDA.

Between v0.8.0 and v0.9.0, it's between 30% and 120% slower. on an RTX 3090 :(

cargo bench --features=cuda
gpu_calc_jones          time:   [1.8791 ms 1.8798 ms 1.8803 ms]
                        change: [+36.252% +36.352% +36.465%] (p = 0.00 < 0.05)
                        Performance has regressed.
gpu_calc_jones 100000 dirs
                        time:   [71.673 ms 71.845 ms 72.039 ms]
                        change: [+117.13% +117.92% +118.71%] (p = 0.00 < 0.05)
                        Performance has regressed.

[d3v-null]

reverting back to stack allocation fixes the regression, segfault seems to not be reappearing?

cargo bench --features=cuda
gpu_calc_jones          time:   [1.8807 ms 1.8813 ms 1.8817 ms]
gpu_calc_jones 100000 dirs
                        time:   [39.354 ms 39.412 ms 39.481 ms]