librbd/cache/pwl: avoid inconsistencies in ImageCacheState

librbd/cache/pwl: avoid inconsistencies in ImageCacheState

When empty and/or clean bools are updated in I/O handling code paths,
ImageCacheState becomes inconistent for a short while: e.g. with clean
transitioned to true, dirty_bytes counter could still be positive
because the counters are updated only in periodic_stats().  Move to
updating the counters in update_image_cache_state(Context*) to avoid
this.

update_image_cache_state(Context*) now requires m_lock -- most call
sites already hold it anyway.  The only problematic call site was
AbstractWriteLog::shut_down() callback chain: perf_stop() needed to
be moved to the very end since perf counters must be alive now for
update_image_cache_state() to work.

Don't override expect_op_work_queue() in unit tests: completing
context in the same thread now results in a deadlock on m_lock in
all test cases that call AbstractWriteLog::init().

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
(cherry picked from commit 016882925a63f4f03a9c445d008b2325d479bc30)