Pre-fetching of collection summaries was quite expensive and we did not use
those summaries very often. Removing the cache completely, now we query
summaries each time but only for the collections that are actually used.

`DefaultCollectionManager.refresh` now runs a single query to fetch
full contents of collection_chain table. This removes update logic from
collection record classes which became simple data classes now.

Static tables do not really need schema verification because we rely
on version numbers in butler_attributes. Verification and reflection
may still be usefule for dynamic tables (tags/calibs) but we now delay
it until the tables are actually used.

Adds special cache classes for collection and summary records and
an additional structure that holds caches. New registry method is
a context manager that enables caches temporarily for the duration
of that context.

Unlike collection caches, dataset type cache is always on, this
helps to reduce number of queries in `pipetask run` without the need
to explicitly enable caching in multiple places.

Registry shiim now redirects to this method instead of Registry
method. RemoteButler raises NotImplementedError but may do something
non-trivial later when we know how caching is going to work with
client/server.