rbd: only get snap context for write requests

rbd: only get snap context for write requests

Right now we get and release the header semaphore every time we
process a request for an rbd image.  We do this because for write
requests we need to supply the snapshot context, and we can't
safely get a reference to it without holding that semaphore.

There's no need to get the snap context if we're doing a read,
so avoid doing so in that case.

The rbd_device->exists field can be updated asynchronously, changing
from set to clear if a mapped snapshot disappears from the base
image's snapshot context.  Although I don't think synchronizing
access to this carefully is that critical, it is converted here to
be an atomic variable so a request is aware the flag is clear as
soon as that is known.

Signed-off-by: Alex Elder <elder@inktank.com>