Exploit additional concurrency in the training algorithm

[rouson]

This PR switches two do loops in the training algorithm to do concurrent constructs. Inspection of the code indicates that there are no values defined in one iteration that are accessed in a subsequent iteration. This conclusion was also tested by temporarily changing the stride of each affected do loops from the default (1) to -1 in order to reverse the order of the iterations. With a stride of -1, all tests pass and app/train-cloud-microphysics.f90 yields the same convergence behavior as with a stride of 1 when run as follows using a training data set from ICAR.

./build/run-fpm.sh run -- --base training --start 720 --epochs 40 

[everythingfunctional]

LGTM

[rouson]

Thanks for the review, @everythingfunctional !