Summary

This PR expands the capabilities of amrex::Gpu::Buffer<T> so that it can be first default constructed, then modified by the CPU, and lastly copied over to the GPU, instead of everything happening in the constructor.

Example:

amrex::Gpu::Buffer<int> buf;

buf.resize(n);

for (int i=0; i<n; ++i) {
    buf[i] = i*i;
}

buf.copyToDeviceAsync();

int * ptr = buf.data();
// Use ptr inside ParallelFor

// optional:
// Change values of ptr inside ParallelFor
buf.copyToHost();
// Use buf.hostData() or buf[] on the CPU

Additional background

Follow-up to discussion in #4640

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