* add gpu_cpu_map_matmul_to_mul_pass to kGpuLowerPrecisionPasses

* disable fc_elementwise_layernorm_fuse_pass in mixed precision