Disappearing (validate) constraint

[yevhenii-nadtochii]

This PR showcases unexpected behavior of codegen validation in core-java. Please note, it is based on class-based-routing branch.

For some reason, when (if_invalid) is applied for Subscription.topic field, its generated code doesn't contain the (validate) check in validate() method.

In StandTest, the modified line should have contained buildPartial() from the very beginning.

Take a look at comparison of the generated validate() method with and without (if_invalid) option:

Discovered in no longer needed #1563.

[yevhenii-nadtochii]

@alexander-yevsyukov PTAL

I mentioned this question in the morning.

[yevhenii-nadtochii]

@alexander-yevsyukov I've reproduced it in validation.

Caused by the use of the deprecated (if_invalid).msg_format instead of (if_invalid).error_msg. Hard to notice. As a result, the error message remained empty, whereas (if_invalid) was discovered by the corresponding projection and expected a non-empty message in error_msg. The validation library throws an exception in this case.

There are the following problems:

1. It looks like we are going to keep msg_format for first releases of V2, so it either should be a somehow workable option with a warning, or an explicit compilation error.
2. Runtime exceptions from validation are hidden, but they are actually compilation errors.
3. Deprecation warnings from protoc are hidden, if they exist at all.