This commit defines new trainable_engine_t components that store
the padded arrays used in the new deep network training algorithm
to store the weights (w), biaes (b), and layer widths (n).  These
components are redundant: the same information stored in updadded
arrays in the parent inference_engine_t type.  If it turns out
that inference performs comparably or better than with the
unpadded data structures -- despite the indirect addressing
required with the padded forms -- then trainable_engine_t can be
refactored to no longer extend inference_engine_t.

This commit switches the AND-gate test to use the recently defined
trainable_engine_t padded-array components for inference.

This commit sets the number of training iterations in each unit
test so that the tests pass repeatedly even when the random_init
subroutine's "repeatable" argument is set to .false. in test/main.f90.