Use linear builder-style compilation of EExpression

Current compilation of EExpression that relies on following pattern:

vm::CodeFragment compileDirect(CompileCtx& ctx) {

  vm::CodeFragment code;

  code.append ...

  return code;

}

this results in O(N*D) complexity, where N is size of expression and D is depth of tree.

Instead, use linear compilation that uses the following pattern:

struct CompileCtx {

    vm::CodeFragment* code;

}; 

void compileDirect(CompileCtx& ctx) {

 ctx.code->append ...

}

in most cases to achieve O(N).

In complex cases, the compileDirect may still allowed to use separate vm::CodeFragments if needed.