[FIX] logging PostgreSQLHandler: expand message + use contextmanagers to handle exceptions and cursor

Before this commit, @mode=primary would be sorta-ignored[0] if the current
view and its parent had the same model: the current view would *still* get
applied (as an extension) when asking OpenERP for its parent. This commit
makes mode=primary views behave regularly, they are *never* applied when
asking for their parent, only when asking for them or their children.

This allows "forking" views, and using extended views in some contexts without
breaking or duplicating the original view

[0] there was actually a problem when asking for the current view directly,
    first its parent would be resolved by applying it, then it would be
    applied to resolve itself, the view would thus get applied twice (oops)

not sure that's actually useful, and can still go always -> enabled -> disabled...

should probably validate that there's an inherit_id (?)

Not used yet, only defined its relation to inherit_id:

not inherit_id + primary -> ok
not inherit_id + extension -> error
inherit_id + primary -> ok
inherit_id + extension -> ok

handle_exception() is supposed to try handling an exception and if it cannot,
re-raise it. Overridden methods must therefore call super() within a try/except
block, and only attempt to handle the exception if super() raised.

[REVERT] b6a7402f, pagereset seems to be useful in some cases. Need to check deeper

1st issue:
When an exception was raised, it was badly handled by the website in case of
website_enabled key. The response page was generated without calling super.
The WebRequest object being responsible to rollback the transaction in case
of errors.

2sd issue:
The _failed attribute is required to rollback the transaction in an WebRequest
object. Previously it was only set in the JsonRequest object (which inherit
from WebRequest), replace by call to super. The attribute _failed is now set
in the WebRequest object.

[FIX] ir_actions: adapt render_report when called in a tests/qweb report context to generate the pdf only if a directory has been provided

opt out with query parameter ?noredirect=1

function fields for other records in the same model. Previously all function fields in the
current model were not computed for some reason not provided by the history.

We therefore compute effective store_ids on which the various trigerred
function fields will be computed again. Those ids are the ids given
in the store_get storage variable minus the deleted ones.

Server-side, view extension is done via xpath. This includes "template" views
full of HTML.

HTML elements often have a bunch of classes, sometimes even semantic
(!). XPath is generally great, but specifically lousy at dealing with
space-separated values: in standard XPath 1.0 to know if an element has a
class 'foo' the predicate is:

    contains(concat(' ', normalize-space(@class), ' '), ' foo ')

and this has to be fully duplicated if there's a second class involved.

Things are slightly better with EXSLT/XPath 2.0 and tokenize, but still not
great:

    tokenize(@class, '\s+') = 'foo'

and the equality check is very weird when unaware of XPath's evaluation rules.

``hasclass`` makes this much simpler to deal with: to get any ``foo`` node
with the class ``bar`` is as simple as:

    //foo[hasclass('bar')

and it can take multiple class, as with e.g. jquery it will return elements
with all specified classes.

Beware though, the predicate function will be called once for each element to
check, since it's implemented in pure python and not profiled elements should
be filtered as much as possible before this point.