[FP16][IE] exclude MVN and NormalizeL2 from precision sensitive marking (#16953)

* exclude MVN from mixed infer

* fix align_mixed_fp32_fp16_types_test.cpp

* fix unit-tests for convert_precision.cpp

* code style fix

src/plugins/intel_gpu/src/plugin/transformations_pipeline.cpp

@@ -194,6 +194,8 @@ void TransformationsPipeline::apply(std::shared_ptr<ov::Model> func) {
 
         // fuse softmax patterns so that they will not be marked as precision sensitive in ConvertPrecision
         manager.register_pass<ov::pass::SoftmaxFusion>();
+        // decompose MVNs that sre not supported in GPU, so the they will be marked as precision sensitive in ConvertPrecision
+        manager.register_pass<ov::pass::MVN6Decomposition>();
 
         //  call ConvertPrecision with keep_precision_sensitive_in_fp32 = true
         manager.register_pass<ov::pass::ConvertPrecision>(fp_convert_precision_map, empty_fuse_map, true);