Intercube allreduce: pe0 cube-mesh reduce + multi-SIP ring/torus/mesh

New intercube allreduce kernel replacing the old flat ring algorithms.
Reduces across the 4x4 cube mesh within each SIP (pe0-only, same-lane),
then inter-SIP exchange on root cube, then broadcast back. Supports
ring_1d, torus_2d, and mesh_2d_no_wrap SIP topologies driven by
topology.yaml. Integrated with dist.init_process_group / dist.all_reduce.

New files:
- src/kernbench/ccl/algorithms/intercube_allreduce.py (kernel)
- src/kernbench/ccl/sfr_config.py (configure_sfr_intercube_multisip)
- tests/test_allreduce_multidevice.py (config-driven, 3 topologies)
- tests/test_distributed_intercube_allreduce.py (full distributed path)
- tests/test_intercube_sfr_config.py (SFR wiring verification)

Modified:
- distributed.py: AhbmCCLBackend uses configure_sfr_intercube_multisip
- topologies.py: added torus_2d, mesh_2d_no_wrap
- install.py: global_E/W/N/S in _OPPOSITE_DIR
- topology.yaml: added system.sips.topology
- ccl.yaml: single intercube_allreduce algorithm
- benches/ccl_allreduce.py: row_wise cube-mesh tensor layout

Removed old flat-ring algorithms and their tests.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

src/kernbench/ccl/sfr_config.py

@@ -0,0 +1,104 @@
+"""SFR configuration for intercube + inter-SIP IPCQ wiring.
+
+Provides ``configure_sfr_intercube_multisip`` which programs PE_IPCQ
+neighbor tables for:
+
+  1. Intercube within each SIP — pe0 of every cube connects to pe0 of
+     its N/S/E/W mesh neighbors (no wrap-around).
+  2. Inter-SIP on ALL cubes — pe0 of cube_c on sip_A connects to pe0 of
+     cube_c on each peer SIP, using ``global_E``/``global_W`` (ring) or
+     ``global_N``/``global_S``/``global_E``/``global_W`` (mesh/torus)
+     direction labels.  Wiring all cubes allows the kernel to
+     dynamically elect the root cube at runtime.
+
+SIP-level topology is read from ``topology.yaml`` →
+``system.sips.topology`` (e.g. ``ring_1d``, ``mesh_2d``).
+Intercube mesh dimensions come from ``sip.cube_mesh.w/h``.
+
+Internally delegates to ``install_ipcq`` with a computed ``rank_to_pe``
+(pe0-only) and a closure-captured ``neighbors()`` function.
+"""
+from __future__ import annotations
+
+import types
+from typing import Any
+
+from kernbench.ccl.install import install_ipcq
+from kernbench.ccl.topologies import _BUILTIN as _TOPO_BUILTINS
+
+
+def configure_sfr_intercube_multisip(
+    engine: Any,
+    spec: dict,
+    cfg: dict,
+) -> dict[str, Any]:
+    """Wire IPCQ for intercube (pe0, mesh) + inter-SIP (pe0, all cubes).
+
+    Args:
+        engine: GraphEngine with ``_components``.
+        spec: topology spec dict (from topology.yaml).
+        cfg: merged algorithm config (from ``resolve_algorithm_config``).
+
+    Returns:
+        The install plan dict from ``install_ipcq``.
+    """
+    cm = spec["sip"]["cube_mesh"]
+    mesh_w = int(cm["w"])
+    mesh_h = int(cm["h"])
+    n_cubes = mesh_w * mesh_h
+    n_sips = int(spec.get("system", {}).get("sips", {}).get("count", 1))
+    sip_topology = str(
+        spec.get("system", {}).get("sips", {}).get("topology", "ring_1d")
+    )
+
+    if sip_topology not in _TOPO_BUILTINS:
+        raise ValueError(
+            f"Unknown sip topology '{sip_topology}'. "
+            f"Available: {list(_TOPO_BUILTINS)}"
+        )
+    sip_topo_fn = _TOPO_BUILTINS[sip_topology]
+
+    world_size = n_sips * n_cubes
+    pe_idx_to_pe: list[tuple[int, int, int]] = [
+        (sip, cube, 0)
+        for sip in range(n_sips)
+        for cube in range(n_cubes)
+    ]
+
+    def _neighbors(pe_idx: int, ws: int, _base: dict) -> dict[str, int]:
+        sip = pe_idx // n_cubes
+        cube = pe_idx % n_cubes
+        row = cube // mesh_w
+        col = cube % mesh_w
+
+        nbrs: dict[str, int] = {}
+
+        # Intercube within SIP (mesh, no wrap-around)
+        if col < mesh_w - 1:
+            nbrs["E"] = sip * n_cubes + (row * mesh_w + col + 1)
+        if col > 0:
+            nbrs["W"] = sip * n_cubes + (row * mesh_w + col - 1)
+        if row < mesh_h - 1:
+            nbrs["S"] = sip * n_cubes + ((row + 1) * mesh_w + col)
+        if row > 0:
+            nbrs["N"] = sip * n_cubes + ((row - 1) * mesh_w + col)
+
+        # Inter-SIP on ALL cubes
+        if n_sips > 1:
+            sip_nbrs = sip_topo_fn(sip, n_sips)
+            for d, peer_sip in sip_nbrs.items():
+                nbrs[f"global_{d}"] = peer_sip * n_cubes + cube
+
+        return nbrs
+
+    mock_module = types.SimpleNamespace(neighbors=_neighbors)
+
+    cfg_copy = dict(cfg)
+    cfg_copy["world_size"] = world_size
+    cfg_copy["topology"] = "none"
+
+    return install_ipcq(
+        engine, spec, cfg_copy,
+        algo_module=mock_module,
+        rank_to_pe=pe_idx_to_pe,
+    )