Clean up KerasImageFileEstimator params

[smurching]

Before releasing DLP 1.0, we should take the opportunity to clean-up KerasImageFileEstimator's params. Specifically we should:

• Update KerasImageFileEstimator.validateParams to check that the imageLoader and modelFile params aren't being tuned. Right now if a user passes said params in a paramMap passed to fit or fitMultiple, we silently ignore the passed params & just use the param values already set in the estimator.
• Check to make sure we don't have any other untunable params like those mentioned above.
• Remove unused params e.g. inputImageNodeName, outputImageNodeName

[yogeshg]

https://github.com/databricks/spark-deep-learning/pull/116 and https://github.com/databricks/spark-deep-learning/pull/117 fixes this

[smurching]

Thanks @yogeshg for the PRs! Closing this issue :)