* Add fused conv support

* Set all saved_model values to non trainable

* Fix TFLite fp16 model export

* Fix int8 TFLite conversion