Add new env-var TF_FP16_CONV_MODE.

The env-var is used to decide how to do internal computation for cudnn convolution when input data type is float16.
For ACCURATE mode, we always use float32 as the internal compute type;
For FAST mode, we include both float16 and float32 internal compute type into auto-tune to pick whichever runs faster.

PiperOrigin-RevId: 174495814