Missing custom input quantization

Currently implemented quantization aware training algorithm gives the chance to test different types of quantization on a model by implementing a custom Quantizer and annotating layers with custom QuantizeConfig classes, but this capability is highly limited by the fact that a default 8bit quantization layer is added after the input layer of the model.
This extra layer will always quantize inputs to 8bits integer, thus it is not possible to explore other types of quantization which could require a different input format.

The quantize_model and quantize_apply methods should provide a more customizable interface that exposes the possibility to implement a user-defined input quantization which will replace the standard 8bit quantization layer. With such an interface it would be possible to enhance the design space exploration capabilities of the QAT algorithm.

System information

• TensorFlow version: 2.5.0-dev20201104
• TensorFlow Model Optimization version: 0.5.0
• Are you willing to contribute it: Yes

The issue names quantize_model and quantize_apply as entry points; read their implementations and trace where the default 8-bit input quantization layer is inserted. Done means these methods expose a user-defined input quantization path that can replace that layer while preserving existing behavior for standard models.