[MXNET-687] Fix spatial transformer op

[anirudh2290]

Description

Spatial transformer op produced incorrect results with USE_CUDA=ON USE_CUDNN=OFF till now. This was due to separate threads writing to the same global memory locations causing undefined behavior.
Replaced all such writes with atomicAdd. I have also reenabled the disabled test: test_stn.
Removed assertion_error param for assert_raises_cudnn_disabled.

This fixes #11568

@tornadomeet @marcoabreu

Checklist

Essentials

Please feel free to remove inapplicable items for your PR.

• The PR title starts with [MXNET-$JIRA_ID], where $JIRA_ID refers to the relevant JIRA issue created (except PRs with tiny changes)
• Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage:
• To the my best knowledge, examples are either not affected by this change, or have been fixed to be compatible with this change

[tornadomeet]

@anirudh2290 the change of spatial_transformer.cu LGTM. the other part @szha

[szha]

LGTM. Looks like CI caught another flaky test, so I just triggered the build again to get past it.

How does the performance (speed) change?

[anirudh2290]

I haven't run any performance tests. Having said that, adding atomicAdd will make it slower but it is required for correctness of the OP.