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.

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.