Fix offset bug in matrix multiply cascade

[hunhoffe]

I had introduced a bug in both matrix multiply whole_array and matrix multiply cascade. In a previous PR, I fixed it for whole_array but did not realize the bug affected cascade as well until it occurred to me to double check recently.

[hunhoffe]

I believe the change I committed fixes the bug, but am running the whole matrix multiply sweep.sh on cascade locally to be more sure before I set this PR as ready for review. This may be related to #1751 but I did not verify (although I do know the bug can present with that type of error). This bug would not have affected matrix_vector, as matrix_vector does not contain any object_fifo_links with offsets, so at most it partially addresses the issue.

At the least, the following case failed before this commit but succeeds after this commit:

M?=640
K?=896
N?=768
m?=16
k?=32
n?=48

[hunhoffe]

I tried config:

M?=32
K?=512
N?=512
m?=32
k?=64
n?=32

and it passed. This may address (one of) the problems raised in the (possibly related) issue (namely, M = 32, n = 32).