Make saving processor and gradient store optional

[luciaquirke]

This is the first of two planned PRs to add support for on-the-fly queries.

Other fixes in this PR:

• For a sharded index build, accumulate one preconditioner over the shards rather than one preconditioner per shard. Because of our disk IO patterns we may still end up with one preconditioner shard for each file but the final one will have the accumulated information from every shard. Can clean this up more later.
• Fix the default value for head_cfg - was a dict (!)
• Add QueryConfig for planned on-the-fly query implementation
• Add test_build file, as we support other saving options (e.g. in-memory return) or on-the-fly queries we can add tests here

Looking for review:

• Made the GradientProcessor's preconditioner dict mutable
• Not sure whether to pipe the QueryConfig into the current script as an optional second variable or duplicate the build.py code into a query.py script etc. to prevent build.py from growing in complexity too much.
  • or secret third option, look for some existing logic to extract out into another file to make build.py simpler