Allreduce sweep: parametrized + xdist parallelism + topology diagram

Refactor the latency sweep from one giant test into 36 parametrized
cases that run in parallel under xdist (~6-8x faster: 1:49 instead of
~10 min). Each case writes a JSON row to a staging dir; conftest
sessionfinish hook aggregates rows on the controller node into
summary.csv and the per-topology + overview plots.

Aggregator gains a CSV fallback so plot-only tweaks no longer require
re-running the sweep.

Overview plot updates:
- 96 KB explicit x-axis marker with vertical dotted line
- horizontal theoretical 2D-torus reference (10600 ns)
- annotation showing both theoretical and simulated values at 96 KB
- drop overlapping 128 KB tick

New topology.png: 2x2 panel diagram showing device-level topology
(ring, torus 2x3, mesh 2x3) and the cube-level reduction inside SIP 0.
Wrap arrows anchor on box edges and arc outside rows/columns so they
do not overlap any SIP.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

tests/conftest.py

@@ -7,11 +7,45 @@ stateful/SimPy-event-consuming and MUST NOT be shared).
 """
 from __future__ import annotations
 
+import os
+
 import pytest
 
 from kernbench.topology.builder import resolve_topology
 
 
+def pytest_sessionfinish(session, exitstatus):
+    """Aggregate parametrized sweep rows into combined CSV + PNG plots.
+
+    Runs on the controller node only (xdist worker processes set
+    ``PYTEST_XDIST_WORKER``; we skip those). Idempotent — does nothing
+    if no sweep rows are present (e.g., when the sweep was filtered out).
+    """
+    if os.environ.get("PYTEST_XDIST_WORKER"):
+        return
+    import importlib.util
+    import sys
+    from pathlib import Path
+
+    mod_path = Path(__file__).parent / "test_allreduce_multidevice.py"
+    if not mod_path.exists():
+        return
+    spec = importlib.util.spec_from_file_location(
+        "_test_allreduce_multidevice_for_aggregate", mod_path,
+    )
+    if spec is None or spec.loader is None:
+        return
+    mod = importlib.util.module_from_spec(spec)
+    sys.modules[spec.name] = mod
+    try:
+        spec.loader.exec_module(mod)
+        agg = getattr(mod, "_aggregate_sweep_plots", None)
+        if agg is not None:
+            agg()
+    except Exception as e:
+        print(f"[conftest] sweep aggregation failed: {e}")
+
+
 @pytest.fixture(scope="session")
 def topology():
     """Session-scoped parsed topology (immutable graph + spec).