hitting the integer limit for MPI-I/O on fat and dense GPU machines

[kostrzewa]

On AMD MI250X (64 GB of GPU memory per GCD), the MG for an 80c160 lattice comfortably fits on two nodes with four GPUs each (eight GCDs each). This means that the 44 GB gauge configuration has to be read by 16 MPI tasks. This means that every MPI task reads around 2.7 GB (2949120086 bytes), which exceeds the silly 2^31 = 2147483648 byte limit of MPI I/O.

This means that we need to split reads and writes exceeding this limit into multiple I/O operations.