Reverting numpy 1.21.2 to 1.19.2 and adding regression

[NikhilRaverkar]

Issue #, if available:

• fixing Update of numpy to version 1.21.0 causing conflict with pyarrow version sagemaker-scikit-learn-container#106 by reverting numpy version.
• Added regression to verify the fix and catch the re occurrence of the issue.
• Although the issue was reported in scikitlearn container, XGBoost uses same version of PyArrow. SO this is just a precautionary change.

Description of changes:

• Numpy 1.2.x broke the integration with Pyarrow 1.0. Rolling back numpy version to 1.19.2
• Added regression test

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[mabunday]

Suggested change

	def test_pyarrow_to_parquet_conversion_does_not_throw_exception(self):
	try:
	df = pd.DataFrame({'x': [1,2]})
	df.to_parquet('test.parquet', engine='pyarrow')
	except Exception as e:
	assert False, "conversion to parquet should not fail"
	def test_pyarrow_to_parquet_conversion_regression_issue_106(self):
	df = pd.DataFrame({'x': [1,2]})
	df.to_parquet('test.parquet', engine='pyarrow')

We don't want the exception to be swallowed if it does fail.

[mabunday]

If you want to add an assertion, you check check that the values are equivalent before and after conversion.

[NikhilRaverkar]

On the exception it will fail the test, thus will fail the build as well.

[mabunday]

Yes, but surfacing the actual exception will make the failure obvious and debugging easier instead of swallowing it with a generic "assert False".