Fix StatisticsPanel pickling: serialize metrics by registry name, atomic writes

[zacsims]

Summary

Fixes the long-standing Bug A in the panel/summary infrastructure where summary_create_panel(..., preset=...) and summary_add_metric(...) fail with:

PicklingError: Can't pickle <function cell_counts at 0x...>:
it's not the same object as spatialtissuepy.summary.population.cell_counts

and, under parallel tool calls, a follow-on:

UnpicklingError: Ran out of input

Two symptoms, two independent root causes in the same subsystem.

Root cause 1: MetricInfo.func is not the module attribute of the same qualname

registry.register() stores the raw func on the MetricInfo object but returns a functools.wraps-wrapped function as the module attribute. Pickle resolves callables by __qualname__ at unpickle time, finds the wrapper at e.g. spatialtissuepy.summary.population.cell_counts, compares identity against the stored raw func, mismatch → refuses to serialize.

Fix: MetricInfo.__reduce__ now pickles by name and re-resolves from the live registry via _resolve_metric_for_pickle(name) on load. Built-in metrics register at import, so panels round-trip without user action. Registered custom metrics round-trip as long as the consumer re-registers before loading; if they don't, unpickle raises a descriptive RuntimeError naming the missing metric instead of an identity-mismatch. Inline metrics added via StatisticsPanel.add_custom_function() are per-panel only and correctly refuse to pickle with a TypeError pointing users at register_custom_metric.

Root cause 2: non-atomic writes produce truncated pickles under concurrency

SessionManager.store_panel / store_data used open(path, 'wb'); pickle.dump(...). Two parallel MCP tool calls can open the same file simultaneously — whichever writer finishes second with less data truncates the file the first had already filled, and the next read sees "Ran out of input".

Fix: _atomic_pickle_dump writes to a per-call temp file (pid + thread id + uuid slug, so in-process threading doesn't collide either) and os.replace-s it into place. Readers always see either the previous file or a complete new one — never a torn write. Failed writes clean up the temp file so no partial bytes leak.

Tests

New tests/test_summary_panel_pickle.py, 11 cases:

• All four preset panels (basic, spatial, neighborhood, comprehensive) round-trip through pickle.
• Registered custom metric round-trips.
• Inline custom function refuses to pickle with a clear TypeError.
• Custom metric removed from the registry between dump and load raises RuntimeError naming the missing metric.
• Eight concurrent threads writing different payloads to the same path via _atomic_pickle_dump always leave a valid complete pickle on disk.
• Failed pickle attempt cleans up the temp file.
• End-to-end: SessionManager.store_panel + load_panel round-trips a basic built-in panel.

Test plan

• pytest tests/test_summary_panel_pickle.py — 11/11 pass.
• Full test suite: 273 passed, 2 skipped, 0 regressions.
• Maintainer review of the __reduce__ semantics for MetricInfo and the atomic-write helper.

Scope

No public API changes. MetricInfo's call / repr / field surface is unchanged; SessionManager.store_panel / store_data keep the same signature. The only new public callable is the module-level _resolve_metric_for_pickle (underscore-prefixed, used only as the pickle reconstruction target).