The whole point is to handle libatomic not being present, so we should
not require it

This is breaking legitimate usecases, we'll reimplement it with the
new mountpoint cache if needed
Fix #414

This reverts commit 7ed13f7c.
Since this commit, most windows tests take more than 15 seconds to
execute, which causes a lot of timeouts

We acquire the connection after acquiring the sqlite lock, so we need to
release those in reverse order, as it's done implicitly in the
Context destructor

Acquiring a priority access means that the current thread will have
priority over other threads attempting to acquire a sqlite lock, it
doesn't acquire a lock in itself, so we need to ignore those when
acquiring a sqlite connection handle.

We need to support acquiring context recursively, whether it comes from
a priority context acquired by the application followed by a write
context within the media library, or 2 write contextes because of
different code path leading to a transaction being conditionnaly opened.

Restoring the tasks is likely to cause a synchronous callback from the
parser, which would attempt to re-aquire the lock, while it was already
held.
Since we don't need to have the lock held when accessing the parser, we
can just release it before.

This reverts commit 5a65ab7b.
This causes the tests to fail intermitently since we now base ourselves
on the idle state, but this commit was causing the idle state to flicker
on and off, which creates a lot of small windows for the test to
consider the scan completed whenever the discoverer turns idle right
after scheduling a task for parsing, and before the parser picks up the
task

This was triggering an assert when the context was already acquired,
which is always the case in the context of VLC which acquires priority
context before executing its queries

Changing the idle state is likely to cause a deadlock on startup when
tasks are restored.
parser::restore() is called from MediaLibrary::initialize() with the
medialib lock held, and changing the idle state synchronously from that
thread would cause onIdleChange to be invoked, which will try to lock
the mutex recursively, leading to a deadlock.

And instantiate that template for thread sanitizer

Fix #411

We're already in the utils::file namespace

If date == releaseYear, we return early

Fix #405

So we can commit the table recreation requests but still hold the lock
to be able to clear connections & cached requests once they all become
invalid
refs #405