added a more naive elemwise implementation that seems just as fast as the recusive-call implementation

Fix test_conv_nnet2 as it had all y value to 0! Lowered the tolerance caused by cumulatif rounding error.

Several minor changes to test_conv_nnet2_classif to run in DebugMode

test_elemwise4 passed; modified setting of stride in alloc_contiguous to set stride 0 for dimensions of size 1