Logcontext fixes

Make client_reader support some more read only APIs

Add some measure blocks to persist_events

Fix some random logcontext leaks.

... to help us figure out where 40% of CPU is going

Only get cached state from context in persist_event

We don't want to bother pulling out the current state from the DB since
until we know we have to. Checking the context for state is just an
optimisation.

Fix missing attributes on workers.

This was missed during the transition from attribute to getter for
getting state from context.

Fix EventContext when using workers

We were:
  1. Not correctly setting all attributes
  2. Using defer.inlineCallbacks in a non-generator

Add concept of StatelessContext, take 4.

Refcator EventContext to accept state during init

Announce deleted devices explicitly over federation.