branchmap: Add a test about writing branchmap and aborted transaction It turn out that we update the branchmap on disk whenever we recompute it… including in a middle of a transaction. This means writing the new branchmap too early (the changeset are not visible yet) and worse… it mean writing branchmap for transaction we rollback. so we introduce a test to highlight the issue (and prevent it to regress in the future). Differential Revision: https://phab.mercurial-scm.org/D12126