GPU: Always deallocate gpu_rand_state in Finalize()

[mirenradia]

Summary

This fixes #3383.

Additional background

I noticed this problem when I tried to build an application with multiple Initialize() and Finalize() calls with HIP e.g. something like:

amrex::Initialize(MPI_COMM_WORLD);
<some AMReX code>
amrex::Finalize();
amrex::Initialize(MPI_COMM_WORLD);
<some AMReX code>
amrex::Finalize();

I'm not sure if this officially supported but it works fine with CUDA (and with MPI so long as you only initialize and finalize this once outside of the above code) but not with HIP which fails at the second initialization here with the error message

amrex::Abort::0::CArena::free: unknown pointer !!!
SIGABRT

Checklist

The proposed changes:

• fix a bug or incorrect behavior in AMReX
• add new capabilities to AMReX
• changes answers in the test suite to more than roundoff level
• are likely to significantly affect the results of downstream AMReX users
• include documentation in the code and/or rst files, if appropriate