Prioritised DQN GPU

[CasBex]

Fixes #973. (Has been tested with CUDA.allowscalar(false))

Two remarks

1. Currently, k and p' are returned as CuArrays. This does not seem to hamper the algorithm so I've not reverted them back, but if that is an issue it can be changed.
2. I have not included tests in the form of an experiment but simply modified the existing JuliaRL_PrioritizedDQN_CartPole. I can add that to the experiments but I think that it will just clutter the RLExperiments package.

[codecov]

Codecov Report

Merging #974 (38f16bb) into main (4f414c6) will increase coverage by 3.96%.
The diff coverage is 0.00%.

@@            Coverage Diff            @@
##            main     #974      +/-   ##
=========================================
+ Coverage   9.66%   13.62%   +3.96%     
=========================================
  Files        214      220       +6     
  Lines       7529     7601      +72     
=========================================
+ Hits         728     1036     +308     
+ Misses      6801     6565     -236     

Files Changed	Coverage Δ
...LearningZoo/src/algorithms/dqns/prioritized_dqn.jl	0.00% <0.00%> (ø)

... and 63 files with indirect coverage changes

[HenriDeh]

Can't do simpler than that. If it works then let's keep the send_to_device for all traces.