[mlir][LLVM] Canonicalize llvm.br and add standalone merge pass (#242139)
Add BrOp::canonicalize, merging branches to single-predecessor blocks,
and llvm-merge-branch-blocks, a pass applying only those patterns.
CIR lowering needs block merging before LLVM IR translation so the
optimizer sees a smaller CFG: the equivalent post-translation sweep
shrank pre-opt dumps 11 to 15 percent and saved about 2 seconds of
optimizer time across five translation units. Full canonicalization
would pay for every pattern and fold the whole module, so running the
patterns standalone, with folding off, gets that win without changing
other operations.