Environment variables that affect execution.

OMP_NUM_THREADS

Sets the number of OpenMP threads per MPI rank.
CUDA_VISIBLE_DEVICES (for CUDA)
ROCR_VISIBLE_DEVICES (for HIP/ROCm)

Sets which GPUs are visible to the program. For example, to make only GPU 0 visible:
```
export CUDA_VISIBLE_DEVICES=0
```
SLATE_GPU_AWARE_MPI

Setting to 1 enables use of GPU-aware MPI within SLATE. If the MPI library is not actually GPU-aware, this will cause segfaults.

Example run

Uses 4 MPI ranks, GPU-aware MPI enabled, with 10 OpenMP threads and 1 GPU per MPI rank. (Output abbreviated.)

slate/test> export OMP_NUM_THREADS=10
slate/test> export CUDA_VISIBLE_DEVICES=0
slate/test> export SLATE_GPU_AWARE_MPI=1
slate/test> mpirun -np 4 ./tester --dim 1234 --dim 1000 --origin d --target d gemm
% SLATE version 2022.07.00, id 04de6a43
% input: ./tester --dim 1234 --dim 1000 --origin d --target d gemm
% 2023-06-26 15:41:22, 4 MPI ranks, GPU-aware MPI, 10 OpenMP threads, 1 GPU devices per MPI rank
 
type  origin  target      m       n       k    nb    p    q      error   time (s)  status
   d     dev     dev   1234    1234    1234   384    2    2   2.55e-16     0.0529  pass
   d     dev     dev  10000   10000   10000   384    2    2   2.05e-16      0.712  pass