Symmetric quantization returns nonzero value for 0 on GPU

Describe the bug
In symmetric quantization, after (fake) quantization, 0 should be exactly 0.
But model-optimization uses the fake quantization API designed for asymmetric quantization so that it could return nonzero values under certain conditions.

System information
Ubuntu 16.04
Tesla V100-DGXS-32GB
Tested in container tensorflow/tensorflow:nightly-gpu

TensorFlow version (installed from source or binary):
nightly container tensorflow/tensorflow:nightly-gpu

TensorFlow Model Optimization version (installed from source or binary):
master

Python version:
3.6.9

Describe the expected behavior
0 returns 0 after symmetric quantization

Describe the current behavior
0 returns a tiny floating point number after symmetric quantization

Code to reproduce the issue

import tensorflow as tf
x = tf.constant([11., 9., 0., 12])
with tf.device("/GPU:0"):
  fq_x = tf.quantization.fake_quant_with_min_max_args(x, -10., 10.)
  fq_x_narrow = tf.quantization.fake_quant_with_min_max_args(x, -10., 10., narrow_range=True)
  qdq_x = tf.quantization.quantize_and_dequantize_v2(x, -10., 10., range_given=True)
print(fq_x.numpy())
print(fq_x_narrow.numpy())
print(qdq_x.numpy())
with tf.device("/CPU:0"):
  fq_x = tf.quantization.fake_quant_with_min_max_args(x, -10., 10.)
  fq_x_narrow = tf.quantization.fake_quant_with_min_max_args(x, -10., 10., narrow_range=True)
  qdq_x = tf.quantization.quantize_and_dequantize_v2(x, -10., 10., range_given=True)
print(fq_x.numpy())
print(fq_x_narrow.numpy())
print(qdq_x.numpy())

Result

[ 1.0039216e+01  9.0196075e+00 -2.4586916e-07  1.0039216e+01]
[ 1.0000000e+01  8.9763775e+00 -3.7252903e-08  1.0000000e+01]
[10.         8.9763775  0.        10.       ]
[10.039215  9.019608  0.       10.039215]
[10.         8.9763775  0.        10.       ]
[10.         8.9763775  0.        10.       ]

Screenshots
If applicable, add screenshots to help explain your problem.

Additional context
The problem is model-optimization uses tf.quantization.fake_quant_with_min_max_vars for "symmetric" quantization https://github.com/tensorflow/model-optimization/blob/master/tensorflow_model_optimization/python/core/quantization/keras/quant_ops.py#L87-L113.
fake_quant_with_min_max_vars are not for symmetric quantization. The one designed for symmetric fake quantization is https://www.tensorflow.org/api_docs/python/tf/quantization/quantize_and_dequantize.

The problem seems to affect GPU only. CPU path returns exactly 0.

Start with tensorflow_model_optimization/python/core/quantization/keras/quant_ops.py at the symmetric quantization path around lines 87-113, then compare its fake-quantization API with the quantize_and_dequantize behavior shown in the reproduction. Run the GPU and CPU examples to confirm the discrepancy; done means symmetric quantization returns exactly 0 for an input of 0 on GPU.