[schema2sql] support NOW and TODAY in check constraints Related to #5154406.

[schema2sql] insert some constraints into the backend BoundaryConstraint, IntervalBoundConstraint and StaticVocabularyConstraint can be implemented in SQL with CHECK constraints. Next steps, not part of this changeset: - migrations - removing no longer necessary pre-insertion work on the python side - translating backend exception due to constraint violation into ValidationError Note: this means that these constraints can no longer be ignored by disabling integrity hooks. Related to #5154406

[test] run unittest_syncschema with postgresql Lets us test more things, since we can actually do ALTER TABLE.

[test] run unittest_migractions on postgresql sqlite's lack of support for a bunch of alter table functionality means some thing just don't get tested otherwise.

[devtools] Put pg sockets in /tmp Makes for shorter paths (limited by sockaddr_un). Also use the default port, since we're not listening on TCP sockets.

[devtools] change the way we start/stop postgresql Instead of having the test db handler start a cluster on demand, use the test module's setUp/tearDown callbacks to do it. This allows to have one data directory (and thus cluster) per test module, allowing different test modules to run in parallel, each using its own database cluster whose path is based on the test module.

[devtools] make test db name for postgresql process-specific In case we run two tests concurrently on the same cluster, better not use the same db name for both.

[devtools] extract functions to start/stop a postgresql cluster

[test] don't leave NULL columns around when making an attribute required It doesn't matter on sqlite (it doesn't do ALTER COLUMN), but when running this test on postgresql it fails to add the 'NOT NULL' constraint otherwise.

[server/session] Roll back in connection exit This was changed in commit 8b35a898b334 "[server] remove cnxset tracking, it is now unneeded". When exiting a CubicWeb Connection, the database connection is put back into the common pool. At that point anything done during the lifetime of the CubicWeb Connection needs to be committed or rolled back. Do a proper rollback to finish the transaction, as was happening before the above-mentionned commit. This showed up on testing with postgresql, when a DB schema modification was blocked by an old connection that was still "in transaction".