Check int overflow/ nan action for torch and add tests

[dongruoping]

Proposed change(s)

• Add the int overflow/ nan action checks for torch

• Use int tensor for global_step for torch: the int32 overflow problem doesn't happen in torch since torch use float tensors by default. But there's a subtle bug in this: when setting the global step to a very large number and then call get_current_step() where the step count is converted to int, there's some numerical error and the result will not be as expected, i.e.

policy.set_step(2 ** 31 - 1)
assert policy.get_current_step() == 2 ** 31 - 1  <---this will fail

This could be a potential bug when trying to resume training from a large step in torch.

• Added tests for int overflow check, verified both TF and torch failed on master and passed with fix in this PR

Useful links (Github issues, JIRA tickets, ML-Agents forum threads etc.)

Types of change(s)

• Bug fix
• New feature
• Code refactor
• Breaking change
• Documentation update
• Other (please describe)

Checklist

• Added tests that prove my fix is effective or that my feature works
• Updated the changelog (if applicable)
• Updated the documentation (if applicable)
• Updated the migration guide (if applicable)

Other comments