Some improvements with tflite conversion

[hunsteve]

• adding logistic activation function to tflite_to_tf converter
  • this is a trivial one-liner
• fixed bug with depthwise convolutions and shared filters in tflite_to_tf conversion
  • there is a bug, if a filter is used in more than one depthwise convolutions, it will be processed in-place multiple times. For the second time, this will fail.
  • fixed it with a not-so-nice solution: monkey-patch the filter tensor object, and if it has been processed already, don't try to process it again
  • other convolution conversion methods might be affected as well
• added basic support for conversion of custom ops for tflite_to_tf conversion
  • TensorFlow Lite flatbuffer file has the name of the custom op. This name is now propagated through the tflite to tf_py conversion, through the attribs dict of the CUSTOM op. This way now it is possible to add a custom converter for this op name, so the actual conversion to NNEF could be done.

[CLAassistant]

All committers have signed the CLA.