Use a faster Stack implementation in TCompactProtocol

TCompactProtocol currently uses a java.util.Stack of Shorts internally to manage its field id deltas. In some profiling, this seems to take more runtime than we'd like. It's probably due to the fact that we are boxing and unboxing the primitive shorts whenever we interact with the stack.

Since we have pretty narrow needs for the stack, I think a purpose-built version could perform much better.