transaction aborts to restore them. That's hackish, and fails as soon as
an "innocent" hand comes in and adds a commit in a test.

* save and restore workflow chain for Person at beginning and start of test
* use setChainForPortalTypes instead of manage_changeWorkflows: the
  first call alters the chain for one portal type, whereas the second affects
  ALL portal types, effectively removing all chains from all portal types not
  passed along in props=. It meant that after each manage_changeWorkflows the
  test was running virtually without any workflow but those of Person, causing
  "fake" failures in other, correct, tests.

This eventually fixes the last failures of testERP5Type


git-svn-id: https://svn.erp5.org/repos/public/erp5/trunk@43101 20353a03-c40f-0410-a6d1-a30d3c3de9de