* add sync_batch_norm (support train, infer, and fp32, fp16, and NCHW, NHWC)

* [NPU] Delete debug codes

* [NPU] Remove FP16