[mlir] Cleanup DialectDocGen to check for the dialect early

We only ever generate documentation for one dialect, so there
isn't a good reason to collect every possible dialect entity.

Differential Revision: https://reviews.llvm.org/D135812