Before the run resource estimator for CPU (OpenMP) and GPU (CUDA)

[kaanolgu]

Provide a small, self-contained test that tells user before launching a large case how much memory it will need per GPU for the CUDA backend, and per rank (host RAM) for the OpenMP CPU backend, so they don't queue a job that would end up being OOM

Proposed behaviour

• One command, reading the real input file (grid, BCs, schemes, decomposition).
• Reports the exact allocator footprint for a sweep of device/rank counts.
• Calibrates the harder to predict term by building a couple of small grids and measuring:
  • CUDA: the cuFFT/cuFFTMp + context workspace
  • OpenMP CPU: the 2decomp/FFT plans + per-rank host buffers, against host RAM.
• Flags configurations that exceed device/host memory and prints the minimum device/rank count required.

Todo list:

• CUDA GPU add initial GPU prerun memory estimator #326
• OpenMP CPU backend (host-RAM estimate)
• Shared core (backend agnostic workspace sweep and backend specific overhead )