We don't use tchar APIs anymore and always have UNICODE set.

No need to guess using tchar and it won't work if _UNICODE is not defined.

The use of FromWide and L'\\' already implied it was the case.

When inserting, we just need to ensure the trigger won't modify the
newly inserted row. When deleting, we don't care about the old row since
it's already gone when the trigger fires.
What we always care about is the position of the modified items relative
to the inserted/deleted row. The media ID doesn't matter since a single
media ID might be inserted multiple times in a single playlist

When a folde represent a mountpoint, we can't generate its name upon
insertion, as the mountpoint path is only known at runtime.
Fix #72

Or rather that concurrent thumbnail request won't cause the 2nd thread
to call std::terminate

The windows tests sometimes fail because of a timeout, increase it to
account for the shared runner running the tests on wine, with an
unoptimized build

ref #67

media_id isn't relevent anymore, and when looking up using only
playlist_id, this index will still help there

fix #67

refs #67

This allows media inserted multiple time to be removed/moved one at a
time, while all the media occurences were removed before
refs #67

refs #67

Instead of [1;N]
This is more natural and more in par with most languages & array/list
indexing.
refs #67

Fix #69

Otherwise we end up deadlocking while acquiring the context. Might as
well fail harder & sooner