Refactor split raw data storage into a DataSetInSeparateSqliteDbFile subclass

[astafan8]

Summary

Stacked on top of microsoft#8219 (targets feature/split-raw-data-sqlite, not main). Implements the subclass option from the "class structure" idea discussion: move the split raw data storage behaviour out of the config-flag conditionals inside DataSet into a dedicated DataSetInSeparateSqliteDbFile(DataSet) subclass, selected at the factory boundary.

This is a prototype for review — it implements the architecture I described in the idea thread so it can be evaluated as an isolated diff.

What changes

Generic seam in the base DataSet

• DataSet gains a small set of overridable hooks whose base implementations reproduce today's main-DB behaviour (all no-ops for a plain DataSet):
  • _setup_results_backend_on_load(read_only) / _setup_results_backend_on_new_run() / _setup_results_backend_on_start()
  • class attribute _creates_results_table_in_main_db = True
• The generic "results may live on a separate connection" routing (_data_conn, the get_parameter_data bypass, the add_results background-writer item) stays in the base and only activates once _raw_data_conn is populated. The backend-agnostic parts added earlier (_results_table_exists, tolerant number_of_results/__len__, deferred subscribers) also stay in the base.

The SQLite backend subclass

• DataSetInSeparateSqliteDbFile(DataSet) overrides only the hooks: create/connect the per-dataset <guid>.db file, record raw_data_db_path, flip _creates_results_table_in_main_db = False. All is_raw_data_storage_enabled() conditionals leave DataSet.

Class selection at the factory boundary

• New runs: new_data_set() and Measurement instantiate the subclass when the feature is enabled.
• Existing runs: a new _load_dataset_from_run_id() helper picks the subclass when the run records a raw_data_db_path. Routed through experiment_container.data_set(), extract_runs_into_db, _copy_dataset_as_is, and _get_datasetprotocol_from_guid.

Tradeoff surfaced by this refactor (worth a look)

Moving class selection to a factory means direct DataSet(run_id=...) construction on a split run no longer auto-detects the raw backend — you must go through _load_dataset_from_run_id / load_by_id. In the current PR microsoft#8219 (config-flag design), DataSet.__init__ auto-detects from the DB row, so direct construction "just works" everywhere.

I routed every construction site in the library through the factory, and updated the make_shadow_dataset test helper. But this is a real behavioural difference for any external code that constructs DataSet(run_id=...) directly. A composition/strategy design (a ResultsBackend object chosen inside DataSet.__init__) would keep the subclass-free base and preserve auto-detection — arguably the best of both. I'm flagging this so we can decide between:

1. this subclass approach (clean class hierarchy, but factory-on-load), and
2. a results-backend strategy object (auto-detection preserved, composition over inheritance).

Either way, this cleanly sets up the future non-SQLite backends goal (a later DataSetInSeparateFile base, or additional ResultsBackend implementations).

Verification

• test_raw_data_storage.py (+ new tests asserting new_data_set/load_by_id return the subclass when enabled and a plain DataSet when disabled), test_database_extract_runs.py, test_export_datasets_and_create_metadata_db.py: 64 passed, 1 skipped.
• test_dataset_basic.py, dond/, test_measurement_extensions.py: 333 passed in default mode.
• Full tests/dataset suite in default mode: 1064 passed, 35 skipped — no regressions.
• Validated with split storage globally enabled (temporary local harness, not part of this PR): same expected/inherent failure set as feat: split raw data storage into per-dataset SQLite files microsoft/Qcodes#8219's baseline (disabled-mode tests, low-level queries.get_parameter_data(ds.conn, ...) bypass tests, db-overview/guid-scan tests that inherently assume data in the main DB). One additional order-dependent flake under the randomized harness (test_do1d_additional_setpoints_shape) passes in isolation and in default mode.
• Pyright: 0 errors. Ruff check + format: pass.