[MTD] CFI flash locking reorg for XIP

This reworks the XIP locking to make sure no lock primitive is ever
called from XIP disabled paths even if in theory they should not
cause any reschedule.  Relying on the current spinlock implementation
is rather fragile and not especially clean from an abstraction pov.
The recent RT work makes it even more obvious.

Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>