TFLiteConverter: Support QuantizeAndDequantizeV4

[lgeiger]

This PR adds support for tf.quantization.quantize_and_dequantize_v2 to the TFLite converter.

Under the hood it is implemented as a canonicalizer that converts QuantizeAndDequantizeV2 (used by tf.quantization.quantize_and_dequantize) to QuantizeAndDequantizeV4 (used by tf.quantization.quantize_and_dequantize_v2). This can be done since both ops share the same kernel implementation in the forward pass. Secondly the pattern in the TFLite converter is changed to handle QuantizeAndDequantizeV4 including a unittest that verifies the behaviour.

This PR is a follow up on #47225 /cc @smit-hinsu

[abattery]

@liufengdb @teijeong @jpienaar could you review this PR?

[jpienaar]

LGTM code wise, are there any differences between 2 and 4 representationally?

[lgeiger]

Thanks for the fast review!

are there any differences between 2 and 4 representationally?

Not as far as I can tell. They share the same implementation in the forward pass:

tensorflow/tensorflow/core/kernels/quantize_and_dequantize_op.cc

Lines 419 to 422 in a86ff79

	REGISTER_KERNEL_BUILDER(Name("QuantizeAndDequantizeV4") \
	.Device(DEVICE_CPU) \
	.TypeConstraint<T>("T"), \
	QuantizeAndDequantizeV2Op<CPUDevice, T>); \

so the only difference is in the gradient op which is not relevant in this case.