[V1.x] ONNX support for Softmax -- optimize for axis=-1 case

[Zha0q1]

Before this pr we would manually expand MXNet softmax. This was due to the mismatch between the configuration of mxnet and onnx softmax.

However we can adapt the onnx softmax to behave the same as mxnet softmax for the axis==-1 case. This is exactly how Gluon BERT uses softmax. With this optimization we 1) improve the runtime 2) avoid NaN issue caused by Exp when running inference with TensorRT in fp16 mode.

[mxnet-bot]

Hey @Zha0q1 , Thanks for submitting the PR
All tests are already queued to run once. If tests fail, you can trigger one or more tests again with the following commands:

• To trigger all jobs: @mxnet-bot run ci [all]
• To trigger specific jobs: @mxnet-bot run ci [job1, job2]

---

CI supported jobs: [miscellaneous, unix-cpu, centos-cpu, clang, website, unix-gpu, edge, sanity, windows-gpu, centos-gpu, windows-cpu]

---

Note:
Only following 3 categories can trigger CI :PR Author, MXNet Committer, Jenkins Admin.
All CI tests must pass before the PR can be merged.

[josephevans]

Is there a testcase for axis = -1?
nevermind, I see there already is a test case :)

[waytrue17]

LGTM