Tracker: Refactor IconCache size handling and usages.

* Instead of taking an icon_size, which we were having to cast
random integers to anyway, just take a BSize and convert internally
as needed. This simplifies a lot of usages of IconCache methods.

* Compute what B_MINI_ICON size will be at startup. This way,
we do not wind up caching "mini" icons in the fLarge*Icon variables
under HiDPI.

This does have a downside that if anything actually
does try to fetch "true mini" (16x16) icons when the real
ComposeIconSize(B_MINI_ICON) is larger than that, it will wind up
(confusingly) in fLarge*Icon, but that should not cause problems
and after this commit should not happen at all, anyway.

* Make mini-icon-mode use ComposeSize instead of the hardcoded 16x16,
and adjust metrics computations around it.

* Fetch larger icons in MountMenu logic. Also use BString::SetToFormat.

* Remove an unused, deprecated method from BPoseView.

* Rename variables in thumbnail generation code to match new behavior.

Tracker: Overhaul thumbnail generation logic.

* Use a JobQueue and BJobs to generate the thumbnails, instead of
spawning a potentially unlimited number of threads (which can
of course rapidly exhaust resources.) Use two threads: one for
smaller files, and another for larger files.

* Directly insert the new thumbnails into the icon cache once they
have been generated, avoiding the port-search dance on filesystems
that do not support writing attributes (or at least large ones.)

* Skip calling mimeset, is it not needed after the previous commit.

* Combine all the duplicated image-scaling logic into one function.

May help with or fix #17225, #17619, and other thumbnail-related matters
e.g. #17557. Also addresses comments from the mailing list last summer.

Tracker: Move thumbnail-manipulation code to its own file.

Preparatory work for refactoring thumbnail generation to not use
so many resources (threads, ports, etc.)