Introduce a parallel, async-capable measurement API for QCoDeS that
decouples plan description from execution. Plans are Python generators
yielding typed messages (Set, Read, Sleep, Emit, OpenRun, CloseRun); a
MeasurementEngine drives them on a dedicated worker thread and publishes
events (RunStarted, RowEmitted, RunStopped) to pluggable sinks via a
publisher thread with a bounded queue.

Key properties of the tracer-bullet implementation:

- Non-blocking submission via engine.submit(...) returning a RunHandle
  with futures for the dataset and final result.
- Cancellation via it.throw(CancelRequested) at the next yield point;
  plan try/finally blocks always run for safe cleanup. scan_1d
  guarantees a ramp-to-zero of the swept parameter on any exit path.
- Cancel flag re-checked between .send() and dispatch to prevent one
  extra Set/Read leaking to hardware after cancellation.
- handle.wait() returning guarantees data is durably committed: the
  publisher thread completes futures only after all sinks have
  processed RunStopped.
- SqliteSink reuses existing DataSaver/Measurement infrastructure;
  opens its own SQLite connection on the publisher thread to satisfy
  check_same_thread=True. ~50 lines of integration code.
- MemorySink for tests and lightweight in-memory consumption.

Public surface:

  import qcodes as qc
  ds = qc.measure_v2.scan(LinSweep(g, 0, 1, 11), measure=[i])

Advanced users construct an explicit MeasurementEngine and submit plans
built via run(...) + plan-builders (scan_1d, or custom generators).

Tracer scope is deliberately small: 1D scans of scalar parameters with
the default SQLite sink. Multi-dim scans, Call/Describe messages,
parallel reads via underlying_instrument grouping, queueing, pause/resume,
live-plot/progress sinks, and a dond shim are P0-P2 for the v1 release.
See src/qcodes/measure_v2/DESIGN.md for the full architecture.

50 tests across L1 (message-stream), L2 (engine + MemorySink), and
L3 (engine + SqliteSink) levels. All three end-to-end acceptance
scenarios pass.

Co-authored-by: Copilot <[email protected]>

- Add @overload signatures to qc.measure_v2.scan() so wait=True returns
  DataSetProtocol | None and wait=False returns RunHandle.
- Narrow Future[Any] -> Future[DataSetProtocol | None] for RunHandle.dataset
  and _Submission.dataset_future. Callers no longer need to cast to access
  .cache.
- Type all test fixtures as Generator[T, None, None] (they yield, not return).
- Move type-only application imports (e.g., ParameterBase) into
  TYPE_CHECKING blocks in test files.
- Tutorial notebook: add type annotations throughout the code cells
  (function signatures, key variable annotations), and place type-only
  imports under TYPE_CHECKING. Notebook still executes end-to-end.

pyright clean on src/qcodes/measure_v2 + tests/measure_v2 (0 errors).
All 70 tests pass. prek hooks pass.

Co-authored-by: Copilot <[email protected]>