also str representation of unicode character is different.

Co-authored-by: Carlos Ramos Carreño <carlos.ramos@nexedi.com>
Co-authored-by: Emmeline Vouriot <emmeline.vouriot@nexedi.com>
Co-authored-by: Kazuhiko SHIOZAKI <kazuhiko@nexedi.com>

this works with astroid 3.2.0 and pylint 3.2.0

on python2, we use have a [python patch][patch] to globally make
asyncore use poll, at that time we were using medusa and it was not
exposing an option to use something else than select.

[patch]: https://lab.nexedi.com/nexedi/slapos/-/blob/180f1553fa7774f05ae8e739dfef43fa2ae56c0a/component/python-2.7/asyncore_poll_insteadof_select.patch

also add the missing tests

Because the librairies used here were never ported to python3.

Notable changes:

  - ERP5Site_createGoogleUserToOAuth is dropped
  - internal API changed radically, so customizations made by overriding
    scripts are broken.
  - the core logic is now implemented in a connector class (still in
    portal_oauth for simplicity, but it would be simpler to move it to
    portal_web_services)

No changes required in the google console, the redirect uri is still
ERP5Site_receiveGoogleCallback

page templates unicode resolver already takes care of coercing when
needed and this is directly compatible with python2 and python3
Co-authored-by: Kazuhiko SHIOZAKI <kazuhiko@nexedi.com>

This only works on python2

for same reason as the xhtml_style version, there is no problem with
having more than one listbox, see d2dedca3 (core: remove useless log
in Base_callDialogMethod, 2020-07-03)

The test testCMFActivity.TestCMFActivity.test_onErrorCallback was
failing because some weak references were not dropped.
The reason was that _DequeueMessageException was declared as global
and it kept a traceback including the variables pointed by the weak
references, preventing their garbage collection.
This was fixed by using a class for the _DequeueMessageException
instead of a singleton, and creating a new instance on every raise.

This test assumed that activating a method with a string of length `n+x`
as argument would be serialized in `x` more bytes that activating the
same method with a string of length `n` as argument, which is only true
for protocol 0 for multiple reasons ( short strings and long strings
are serialized differently, "frames" are used with protocol 5) and
db.string_literal seems to introduce a difference.

Change the test to mock Message.dump to produce longer dumps instead of
relying on this assumption.

_resultForDoCleanups no longer exists on python3, getting the status
of a test in teardown is very different accross python versions ( see
https://stackoverflow.com/q/4414234 ) and just executing this code when
the test failed only has the problem that a test method may count as
two failures or errors, which is after all not a big problem.

Manager username and password are now available as class attributes.

Many tests have been updated to not generate manager user and use the
existing one instead. When tests need to create users (for example
users in the root acl_users), we try to delete the users afterward.

Another significant changes is that we no longer have the same user in
ERP5/acl_users and /acl_users, some tests were logging in as the root
user in ways that never seemed intentional.

ERP5TypeFunctionalTestCase was updated to use self.manager_username and
manager_password instead of hardcoding ERP5TypeTestCase: and also some
cleanups, the HTTP server is already started at this point and host/port
were not used.

Some general changes to test framework and shared test code

See merge request nexedi/erp5!1936

this happens with new ghostscript, but this does not seem a critical
problem.

The test was only testing with empty cache

This should not happens, but if this happens we should not just return
in the middle of the loop, because this returned value is not used,
so we would just silently edit partially.