[Bugfix]Revert PR #2420 cuda/nms changes

[Laurawly]

According to discussion forum: https://discuss.tvm.ai/t/mali-cl-out-of-resources-error-for-mxnet-ssd-models/1863. Found the bug caused by PR #2420's cuda/nms.py change. Reverting it back. @apivovarov

[apivovarov]

Can you add unit tests for that fix?

[Laurawly]

Can you add unit tests for that fix?

nms unit tests already exits. The bug only happens for mali rk3399 and I don't think they have it in CI (correct me if I'm wrong.)

[apivovarov]

Can you add unit tests for that fix?

nms unit tests already exits. The bug only happens for mali rk3399 and I don't think they have it in CI (correct me if I'm wrong.)

Can we add unit test for sort_ir method which fails without this fix and works fine with this fix?

[vinx13]

@Laurawly do you have idea why this caused error?

[Laurawly]

Can you add unit tests for that fix?

nms unit tests already exits. The bug only happens for mali rk3399 and I don't think they have it in CI (correct me if I'm wrong.)

Can we add unit test for sort_ir method which fails without this fix and works fine with this fix?

Sort_ir is included in nms op. It's not yet separated as a single op. Again this case will only fail on certain mali GPUs.

[Laurawly]

@Laurawly do you have idea why this caused error?

@vinx13 I'm not quite sure. It seems to me that the less vthread we use, the more workload each vthread has which might cause cl_out_of_resource error in this case. Also I tried adding storage sync for previous method but it still failed.