Enable double caching: CPU cache and GPU cache

[huiyuxie]

Fix #59.

The GPU-only cache strategy causes the indicator functions to run slower on GPU arrays. We are enabling an additional cache on the CPU as a hotfix strategy. However, not all data needs to be stored on both the GPU and CPU, so we should differentiate them in the future.

The latest benchmark results for Euler equations for shock capturing should be given here.

[huiyuxie]

1D benchmark

[ Info: Time for reset_du! and volume_integral! on GPU
BenchmarkTools.Trial: 10000 samples with 1 evaluation.
 Range (min … max):  35.300 μs … 361.000 μs  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     42.200 μs               ┊ GC (median):    0.00%
 Time  (mean ± σ):   52.005 μs ±  22.623 μs  ┊ GC (mean ± σ):  0.00% ± 0.00%

  ▄▆██▇▆▄▃▃▂▁▂▄▄▃▂▃▃▂▂▁▁  ▁▁ ▁▁▁              ▁▁▁ ▁  ▁         ▂
  █████████████████████████████████▇███████▇▆███████████▇▆▄▆▄▅ █
  35.3 μs       Histogram: log(frequency) by time       128 μs <

 Memory estimate: 3.94 KiB, allocs estimate: 40.

2D benchmark

[ Info: Time for reset_du! and volume_integral! on GPU
BenchmarkTools.Trial: 10000 samples with 1 evaluation.
 Range (min … max):  197.100 μs …   5.657 ms  ┊ GC (min … max): 0.00% … 23.73%
 Time  (median):     262.100 μs               ┊ GC (median):    0.00%
 Time  (mean ± σ):   282.205 μs ± 163.339 μs  ┊ GC (mean ± σ):  2.99% ±  5.30%

   ▄▇█▆▆▃▁▂▂▄▄█▇▂
  ▄████████████████▆▆▆▅▅▄▄▄▄▄▄▃▂▂▂▂▂▂▂▁▂▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁ ▃
  197 μs           Histogram: frequency by time          553 μs <

 Memory estimate: 258.50 KiB, allocs estimate: 42.

3D benchmark

[ Info: Time for reset_du! and volume_integral! on GPU
BenchmarkTools.Trial: 3373 samples with 1 evaluation.
 Range (min … max):  1.183 ms …   5.663 ms  ┊ GC (min … max): 0.00% … 67.39%
 Time  (median):     1.434 ms               ┊ GC (median):    0.00%
 Time  (mean ± σ):   1.478 ms ± 296.514 μs  ┊ GC (mean ± σ):  1.62% ±  5.84%

    ▁▅▇██▅▃▂
  ▃▅████████▇▄▄▃▃▃▂▂▂▂▂▂▂▂▂▁▂▁▁▁▂▂▁▁▁▁▁▁▁▁▁▁▁▂▁▁▁▁▁▁▁▁▁▂▁▁▁▂▂ ▃
  1.18 ms         Histogram: frequency by time        3.57 ms <

 Memory estimate: 642.64 KiB, allocs estimate: 43.

They are still larger than what we obtained from the profiling - something other than the GPU kernels (e.g., allocating the CPU arrays) is dominating the overall runtime.