Keep track of stream usage for arrays

[pxl-th]

• Add Managed struct that keeps track of TLS streams and correctly switches between them. This in turn allows us to not use global stream in finalizers.
  Ref: Track CuArray stream usage CUDA.jl#2236
  Allows us to use tls streams in finalizers instead of global one.
• Remove REPL hook.

Performance on Nerf.jl benchmark (1k training steps & 10 rendered 1080p frames):

• 0.8.11:

Training
 38.445143 seconds (3.35 M allocations: 145.687 MiB, 0.15% gc time)
Rendering
141.926297 seconds (26.43 M allocations: 930.844 MiB, 0.17% gc time)

• This PR:

Training
 34.165407 seconds (3.24 M allocations: 153.949 MiB, 0.15% gc time)
Rendering
116.225133 seconds (27.26 M allocations: 1.039 GiB, 0.22% gc time, 0.04% compilation time)

#628 also had an impact on performance.