fix: normalize default partitions in in-memory persisters

[MIRS571]

Follow-up to the non-blocking default-partition suggestion in #958.

The in-memory persisters currently store a checkpoint under the raw None key but return partition_key="" in its metadata. Reloading with that returned key misses the checkpoint. SQLite already normalizes both inputs to the same default partition.

Changes

• Normalize only None to "" in synchronous and asynchronous in-memory save/load/list operations, including stored checkpoint metadata.
• Preserve non-empty keys verbatim and retain non-inserting load behavior.
• Add 16 parameterized regression cases for alias lookups, returned-key reloads, mixed-key histories, duplicate sequence IDs, multiple applications, missing reads, and named-partition isolation. Synchronous lookup cases also run against SQLite as a control.
• Document the compatibility change.

Compatibility

None and "" are now aliases, not two isolated in-memory partitions. Applications using both for the same app_id share a checkpoint history. Existing last-save semantics remain unchanged, including when selecting a specific sequence ID. Callers requiring separate partitions should use distinct non-empty string keys. SQLite behavior is unchanged.

How I tested this

On Python 3.11:

• Before the fix, the initial new regression cases produced 10 failures and 6 passes.
• python -m pytest tests/core/test_persistence.py -q: 66 passed.
• python -m pytest tests/core -q -k "not test_visualize_dot_output": 446 passed, 4 deselected. The complete core run was attempted; those four graph-rendering cases cannot run locally because the Graphviz dot executable is absent.
• Black, isort, flake8, AST, whitespace, EOF and ASF-header pre-commit checks passed on the changed files. The frontend hook was skipped because this change contains no frontend files.
• git diff --check passed.

Notes

The complete repository test suite, other Python versions and the Sphinx documentation build were not run locally.

Checklist

• Informative title and one focused goal.
• Applicable pre-commit checks passed, with the frontend exclusion noted above.
• Changed behavior is covered by regression tests.
• Compatibility behavior and class documentation are updated.