Quite WIP.  Have marked some TODOs.  Will add some provisional changes to
manager.

In testing, I just found a typo I made yesterday.  Also, reopening the
SharedMemory even readonly was failing, so I set the bridge to only open once
and then change view (as it should, to be fair).  Will think about nice
encapsulation of this.

Was trying to map out behavior of DsscGpuRunner as an FSM and found that this
was missing.

- Move shared memory buffer from kernel runner / buffer device to correction
device itself
  - Wrap the ring buffer around shmem for convenience
  - Reused shared memory segment, just change shape of ndarray view of it
- Rename kernel runner (was PyCudaPipeline) to DsscGpuRunner
- Move all gpu interaction, to DsscGpuRunner
  - Explicit load methods (should control more FSM-like)
- Pulse filter removed temporarily
  - Part of "splitter" section, gets in the way of next refactoring step
  - Will add again after switching to correct first, reshape second operation