Fix offset in recv buffer of single precision particle communication (#…

…3896)

A similar bug was fixed in #3758 for the particle send buffer. But we
did not notice the issue in the receive buffer at the time. The offset
for the `i`th buffer might be aligned for the `i`th buffer's
communication data type, not the previous buffer.

The reason the bug only appears for large single precision runs is
because for double precision the particle size is aligned with the size
of unsigned long long used in communication, whereas for single
precision this is not the case.

Src/Particle/AMReX_ParticleCommunication.H

@@ -481,10 +481,10 @@ void communicateParticlesStart (const PC& pc, ParticleCopyPlan& plan, const SndB
     for (int i = 0; i < NProcs; ++i) {
         if (plan.m_rcv_num_particles[i] > 0) {
             RcvProc.push_back(i);
-            rOffset.push_back(TotRcvBytes);
             Long nbytes = plan.m_rcv_num_particles[i]*psize;
             std::size_t acd = ParallelDescriptor::sizeof_selected_comm_data_type(nbytes);
             TotRcvBytes = Long(amrex::aligned_size(acd, TotRcvBytes));
+            rOffset.push_back(TotRcvBytes);
             TotRcvBytes += Long(amrex::aligned_size(acd, nbytes));
             plan.m_rcv_pad_correction_h.push_back(plan.m_rcv_pad_correction_h.back() + nbytes);
         }