[Experimental] Add memory recording which can then be bulk-freed

[pxl-th]

Add memory recording which can then be bulk-freed.
Because GC is not aware of GPU memory and relying on it leads to significantly higher memory usage and big delays, the idead is to record all device-allocations that happen during AMDGPU.record_memory!(true)
and when AMDGPU.record_memory!(false) is set - bulk-free them.

This is useful for example in DL scenarios where you have your model parameters and compute gradients and update them in-place and once you did this, all of the intermediate allocations can be immediately freed.

θ = <parameters>
AMDGPU.record_memory!(true)
∇ = gradient(θ) do θ
   ...
end
apply!(θ, ∇)
AMDGPU.record_memory!(false) # bulk-free all allocations that happened during recording.

Additionally, you can remove array from recording with AMDGPU.remove_record!(x) or disabling recording for that allocation with AMDGPU.record_memory!(false; free=false) (just remember to enable it again).

Example of memory usage with GaussianSplatting.jl:

Before	This PR