This fixes build on Windows with MSVC, mostly linking issues,
template instantiation, silencing overly-pedantic warnings and
fixing a few actual warnings that GCC/Clang don't warn about.

Sandro Knauß authored Nov 02, 2015 and Sandro Knauß committed Mar 21, 2016

Cherrypicked from 1b3d8f37422a3f452e7863a7d9155e4fee727094

Differential Revision: https://git.kolab.org/D67
Differential Revision: https://phabricator.kde.org/D1159

Now that the headers are in the same repo, we need to adjust all the
includes of it and include paths.

Most of the time the changelog members in ItemPrivate are unused and
wasting memory. This change introduces ItemChangeLog singleton where
Items can store their changelog if they need to in a common QHashes.

This save another at least 40 bytes per each Item instance.

All models are in core/models, having some in widgets unless there's a real reason
(dependency-wise) would just be confusing.

Collection and Item now have each their own implementation of attribute handling
(which we can factor out into some shared code later), and by not having to deal
with the polymorphism of Entity, we can get rid of vtable in ItemPrivate and
CollectionPrivate and better re-order member variables. This saves us 16 bytes
per Item and 32 (!) bytes per Collection. Tag now requires 8 bytes more
of memory due to TagPrivate being QSharedData (but we save something on Tag now).

We could also remove the copy-d_ptr-via-temporary-assignment from from Item,
because the usecase only really affects Collections. This should make copying
Items a bit faster (the assignEntityPrivate() workaround was showing up in
callgrind quite high).

The Protocol DataStream uses waitForData() for blocking wait which makes it
unsuitable for use from the main thread. Also the Protocol is now so fast, that
it blocks GUI when opening larger folders simply because we constantly have data
to deserialize and we don't return to event loop often enough. Moving this
to a separate thread is the right thing to do. It also notably speeds up opening
larger folders.

We only needed it for one function which also exists in KCoreAddons.

It's now only used internally in the subscription dialog.

We replaced the usage of SELECT by actually specifying the collection
in each command, so the job is just unnecessary. Now we can get rid
of it on the server and simplify a lot.

Monitor, once subscribing to the NotificationManager starts a new Session using the
identifier received from NotificationManager. The Session is started in the special
NotificationBus mode, so the server will use it to deliver notifications for the given
subscriber through it.

This laos ports all code from NotificationMessageV* to Protocol::ChangeNotification.

Should fix build systems where Boost is not installed in standard prefix.

On case-insensitive systems, #include <exception> includes our own
fancy header "Exception" instead of the stdlib exception header.
To disambiguate this, we resolve a full absolute path to the stdlib
header (through some magic of parsing compiler debug output) and
generate std_exception.h include file that we then use in our
excepton.h to include the stdlib exception header.

It now no longer forces KDELibs4Support and or even QWidgets on its users.

Reverted the change in Akonadi.

This reverts commit 2c547627f02c31cc4fe4387a5f83d2fbc94de317.

This way, adding kde_enable_exceptions is not needed anymore.