osd: dispatch_context inside PG lock

osd: dispatch_context inside PG lock

We have been using PeeringCtx (previously RecoveryCtx) to batch up
query, notify, and info messages and send them to other OSDs.  This is
optimizing for small clusters in a way that doesn't apply to large
clusters and complicates the code.

(FWIW I also don't think the batching was working very well anyway.)

Instead, dispatch everything *inside* the PG lock, and do not accumulate
work across multiple PGs.  Avoid re-using the RecoveryCtx at all by
moving it into the loops.

Signed-off-by: Sage Weil <sage@redhat.com>