Add support for FloatType, IntegerType and LongType inputs to TFTransformer

[smurching]

What

Currently, TFTransformer accepts only DoubleType input columns, casting Double input into the input type expected by the TF graph. This casting implicitly constrains TFTransformer to support only those TensorFlow input types that can be casted-to from double.

This PR removes the cast-from-double operation in TFTransformer, thereby adding support for FloatType, IntegerType and LongType input columns.

Why

The changes in this PR will facilitate adding support for BinaryType input columns in a follow-up

Summary of Changes

• .gitignore: Add files generated by DeepLearningFeaturizer tests and codecov to .gitignore
• tf_tensor.py: Removes extra cast ops added to TF graphs in TFTransformer (this is a breaking change; before, users could feed Doubles into tf.float32, tf.int64, tf.int32 tensors). Instead, determines the type of each input tensor (in _get_placeholder_types) and replaces it with a placeholder of the same type using the TF optimize_for_inference API.
• keras_transformer_test.py: Changed to pass float32 input to Keras graphs (which have float32 input tensors) instead of float64 input.
• tf_transformer_test.py: Modified to test TFTransformer with float, int, and long input columns.

[codecov-io]

Codecov Report

Merging #108 into master will decrease coverage by 0.17%. The diff coverage is 92.85%.

@@            Coverage Diff             @@
##           master     #108      +/-   ##
==========================================
- Coverage   82.84%   82.66%   -0.18%     
==========================================
  Files          34       34              
  Lines        1999     1990       -9     
  Branches       44       44              
==========================================
- Hits         1656     1645      -11     
- Misses        343      345       +2

Impacted Files	Coverage Δ
python/sparkdl/transformers/tf_tensor.py	98.24% <92.85%> (-1.76%)	:arrow_down:
python/sparkdl/graph/utils.py	98.94% <0%> (-1.06%)	:arrow_down:

---

Continue to review full report at Codecov.

Legend - Click here to learn more Δ = absolute <relative> (impact), ø = not affected, ? = missing data Powered by Codecov. Last update e3c5876...5bac9b6. Read the comment docs.

[smurching]

Chatted offline with @sueann - added a warning about the behavior change. Will merge now, thanks @sueann @tomasatdatabricks!