Failed reduction tests for certain numbers of blocks and threads

[adampbeardsley]

In exercise_solutions/cuda/reduction_atomic/kernel.cu, the test fails if we hard-code certain values for the blocks and threads in the kernel launch. For example, making line 112:
sumReduction<<< 32, 32 >>>( size, d_in, d_sum );
results in a test fail, but with no code or runtime error. This fails for certain values of (blocks,threads), but I haven't found a pattern yet. For example,
PASSES:
(31, 31)
(31, 33)
(32, 33)
(33, 31)
(33, 32)
(33, 33)
FAILS:
(31, 32)
(32, 31)
(32, 32)

Some smaller numbers also fail, like (1,2) and (2,1). But (1,1) works.
This is also seen in reduction3 and probably others.