Generalize GROUPED_CONV_2D op to support input product scale > output scale

To be consistent with other convolution ops we have.
TF Lite removed input/output scale restrictions to support additional
use cases for its quantized convolution operations: CONV_2D and
DEPTHWISE_CONV_2D. NN API additionally supports quantized
TRANSPOSE_CONV_2D and GROUPED_CONV_2D. This change relaxes restrictions
on quantized GROUPED_CONV_2D to make it consistent with quantized
CONV_2D, DEPTHWISE_CONV_2D and TRANSPOSE_CONV_2D.
This is a superset of existing GROUPED_CONV_2D functionality.

Bug: 127427474
Test: NeuralNetworksTest_static
Change-Id: Ia2ee0ffae29b7695a1014c8a7131e18a95413307
Merged-In: Ia2ee0ffae29b7695a1014c8a7131e18a95413307
(cherry picked from commit 021cddc5)