[RLlib] Minor fix on json encoding during worker sampling

[sungho-joo]

Why are these changes needed?

This PR fixes a minor bug on json encoding during worker sampling (specifically when output_writer write batch).
Currently, when output_writer write batch, it first transforms data into json file.
If data contains type that is not json serializable, error occurs while dumping.
In rllib, this usually happens in custom env with custom infos.
Luckly, similar issues have been reported before (#2243) and SafeFallbackEncoder has been implemented.

This PR contributes in two way

1. Import and pass SafeFallbackEncoder to json.dumps
2. Improving SafeFallbackEncoder

Regarding the improvement of SafeFallbackEncoder, order of nan checking and np.ndarray type check has been changed.
This is because return of np.isnan(ndarray object) is not boolean but array of boolean.
if np.isnan(np.array([1,1])): -> throws error
Therefore, ndarray type check has to be checked before checking nan
Also, type check of np.bool_ is added.

Using implemented SafeFallbackEncoder,

Related issue number

"Resolves #5785 "

Checks

• I've run scripts/format.sh to lint the changes in this PR.
• I've included any doc changes needed for https://docs.ray.io/en/master/.
• I've made sure the tests are passing. Note that there might be a few flaky tests, see the recent failures at https://flakey-tests.ray.io/
• Testing Strategy
  • Unit tests
  • Release tests
  • This PR is not tested :(

[amogkam]

LGTM, thanks a lot for the patch :)