With previous algorithm, work was given to additional test nodes only when:
- we were previously below the needed capacity
- when another test node was dying

Now, as soon as a new test node is added, we move work of overloaded test nodes to
idle test nodes. We try to move only test suite using many test nodes to avoid having to
wait for building time.

This allows to have better distribution of the work with the idea to have more quickly test results.
This will avoid cases where we have several testnodes assigned to no work at all.

Finally, fixed distribution algorithm to avoid some unfair cases where a test suite might
have more test node than another while they both ask for the same number of test nodes.

Documentation says this step is done in one transaction, so implement
this as documented.

Detecting developer mistake and automagically running this step in smaller
transactions if more than 100 documents seems unnecessary. At least, it
contredicts documentation.

now that we use alarm sense method, another `notify` call is activated

This should fix some race conditions which made post-upgrade steps not
executed sometimes.

* each steps use an defaultActivateParameterDict so that activity they
  may create inherits the same tag, so that next step also runs after
  this). This makes it possible for example to use post-upgrade step on
  documents newly installed in upgrade step.
* use tag / after tag between upgrader steps instead of checking if
  previous alarm had errors. We also now allow force running only one
  step regardless of the previous step completion state, there was no
  strong reason to prevernt this.
* do not use two different active sense and fix method for alarms, just
  use an active sense method understanding `fixit` parameter

It allows idempotent result on any test node

replace method_name:method parameter instead of replacing form action, otherwise method_name like listbox_setPage does not work.

unescape_chr is executed with such string: "C:\Program Files\1234.jpg".
Prevent failing, and return original string in such case.

Previous behavior was to generate reference this way : `sanitized title` if reference was not already existing else `random "-" sanitized title`.
Now it generates `sanitized title "-" random` with:
- `random` at the end to maximize search engine optimization;
- `random` every time present to prevent creating document with the same reference by mistake.

+ update tests

added possibly by mistake.

see 1cb54350

We now use URL only, old address + port way is obsolete.

host and port configuration is obsoleted by URL configuration

Build a conversion_server_url from conversion_server_(hostname/port), so
that we can already stop using the old parameters

The ImportFixer AST processor wasn't correctly handling the alias
for imports. Now it takes them into consideration correctly. A test
to cover this case was added.

The old hack used to catch some exception is not required anymore.
The bug is gone with upstream updates.

So that a jump to the attachment file can be performed

/reviewed-on nexedi/erp5!158

This is a duplicate of !150, because by unknown reasons it's showing problem to merge.

Before, we we're redirecting `sys.stdout` and this doesn't play nice with the
distribute architecture of ERP5 and our Jupyter kernel needs to be adjusted for this.

So, we're now using an AST processor to fix print calls. It will modify the print
and make it write to a different file-like object. All the writes are collected
after code execution and sent to Jupyter.

It's still necesasry though to fix print inside other libraries. But for this
deeper investigation is necessary because we cannot replace print as a statement
inside `exec` contetx, it needs to be used as a function. Code can be compiled to
run with `print` as a function, but then external libraries calls will be
broken.

/reviewed-on nexedi/erp5!157