Add fp16 support to color distortion ops for better performance on GPU. These

ops are typically used for image preprocessing in data input pipeline. When the
input pipeline gets complicated and becomes the bottleneck (for example, in
the SSD (Single Shot multibox Detector) model), offloading partial image
preprocessing to accelerators can be useful to alleviate CPU workload. Adding
fp16 support can further accelerate the processing on GPU and allow the model to
have larger batch sizes.

The ops that are affected in this CL are:
  tf.image.random_brightness
  tf.image.random_contrast
  tf.image.random_saturation
  tf.image.random_hue
  tf.image.adjust_brightness
  tf.image.adjust_contrast
  tf.image.adjust_saturation
  tf.image.adjust_hue

PiperOrigin-RevId: 229422273