* dnn(ocl4dnn): fix LRN layer accuracy problems

- FP16 intermediate computation is not accurate and may provide NaN values

* dnn(test): update tolerance for FP16