use new rand functionality from CUDA.jl

[simsurace]

No description provided.

[simsurace]

These are the benchmark results from the automatic builds:

rand(Float32) in Julia 1.6.0-rc3 using CUDA.jl#tb/speedup_rand:

Benchmarking constant parameter array: should run in less than 2ms on an RTX20xx card
BenchmarkTools.Trial: 
  memory estimate:  704 bytes
  allocs estimate:  21
  --------------
  minimum time:     605.290 μs (0.00% GC)
  median time:      641.916 μs (0.00% GC)
  mean time:        649.830 μs (0.47% GC)
  maximum time:     3.488 ms (0.00% GC)
  --------------
  samples:          7693
  evals/sample:     1

Benchmarking full parameter array: should run in less than 2ms on an RTX20xx card
BenchmarkTools.Trial: 
  memory estimate:  160 bytes
  allocs estimate:  4
  --------------
  minimum time:     706.299 μs (0.00% GC)
  median time:      760.639 μs (0.00% GC)
  mean time:        762.317 μs (0.00% GC)
  maximum time:     5.279 ms (0.00% GC)
  --------------
  samples:          6558
  evals/sample:     1

GPUArrays.gpu_rand in Julia 1.6.0-rc3 with CUDA.jl 2.6.2:

Benchmarking constant parameter array: should run in less than 2ms on an RTX20xx card
BenchmarkTools.Trial: 
  memory estimate:  2.11 KiB
  allocs estimate:  53
  --------------
  minimum time:     902.502 μs (0.00% GC)
  median time:      964.407 μs (0.00% GC)
  mean time:        976.075 μs (0.33% GC)
  maximum time:     3.704 ms (44.78% GC)
  --------------
  samples:          5122
  evals/sample:     1

Benchmarking full parameter array: should run in less than 2ms on an RTX20xx card
BenchmarkTools.Trial: 
  memory estimate:  1.58 KiB
  allocs estimate:  36
  --------------
  minimum time:     1.195 ms (0.00% GC)
  median time:      1.292 ms (0.00% GC)
  mean time:        1.293 ms (0.00% GC)
  maximum time:     1.868 ms (0.00% GC)
  --------------
  samples:          3867
  evals/sample:     1

[simsurace]

Julia version	constant parameters	full parameters
1.5.4	605us	861us
1.6.0-rc2	976us	1293us
1.6.0-rc3 with new rand()	649us	762us