Fake Grammar-Kit dependencies

We use Grammar-Kit to generate Rust parser and PSI classes from BNF-like grammar. Grammar-Kit decides on how to generate PSI based not only on a .bnf grammar file, but also on contents of Java/Kotlin files referenced in that .bnf file. In particular:

1. Grammar-Kit takes into account super classes of a java class referenced using expends property in a BNF rule when deciding on how to generate PSI methods for that BNF rule. If Grammar-Kit finds CompositePsiElement class in super classes, it uses "mixed AST" style, where single class represents both PSI and AST elements (so additional AST node is not allocated). Otherwise, it uses split PSI/AST style where PSI elements wrap AST elements (or stubs).
2. Grammar-Kit takes into account constructor arguments of a java class referenced using mixin or expends property in a BNF rule when deciding on how to generate PSI methods for that BNF rule. In "mixed AST" case, Grammar-Kit just mirrors constructors of the mixin/expends class.

Important point is that Grammar-Kit analyses compiled java classes, i.e. classes referenced in a .bnf file have to be compiled prior to Grammar-Kit invocation. Unfortunately, we can't satisfy this requirement because our fundamental PSI classes are heavily depends on PSI classes generated by Grammar-Kit, so this is a kind of chicken-and-egg problem.

This module solved the problem by introducing essential set of FAKE PSI classes containing only stuff needed to Grammar-Kit. These classes are not included to the final plugin .jar file and furthermore are not used anywhere except in Grammar-Kit.

These hacks will likely become unneeded when this PR to Grammar-Kit will be accepted.