Fixes for @aaroey's review.

Fix MNIST integration test imports and add the data dependencies as a build rule.
GetWeightRange will now only be called in int8mode no calib mode.
Make precision_mode a member of TrtCandidateSelect instead of argument to IsCandidate.
Make precision_mode, use_calibration members of Converter
Remove QINT8 from ConvertDType. Remove QINT8, UINT8, INT8, QUINT8 from RedorderCK. These are no longer needed after removing QuantizeV2 and Dequantize support.
Refactor ApplyQuantizatioNRanges() to make it easier to test.
ConvertQuantize will explicitly look for the different ops by name.
Add checks to Relu6.
Add unit tests for new functions in Converter.
Add comments to clarify some logic