Add SchedulerV2 (pe_accel), DPPolicy overrides, and new benchmarks

- Add cycle-accurate PE accelerator scheduler (SchedulerV2) with tiled
  GEMM/Math pipelines (DMA_IN → GEMM → MATH → DMA_WB)
- Add DPPolicy num_pes/num_cubes/num_sips overrides for single-PE testing
- Support tuple target_pe for targeting specific PE subsets
- Add gemm_single_pe and gpt3_qkv benchmarks
- Switch default topology to pe_scheduler_v2

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

src/kernbench/runtime_api/tensor.py

@@ -105,7 +105,7 @@ class DPMetadata:
     dp_policy: DPPolicy | None = None
     sip: int = 0
     cube: int = 0
-    target_pe: int | str = 0  # int → single PE, "all" → all PEs
+    target_pe: int | tuple[int, ...] | str = 0  # int → single PE, tuple → specific PEs, "all" → all PEs
 
 
 class Tensor:
@@ -166,7 +166,7 @@ class Tensor:
         dp_policy: DPPolicy | None = None,
         sip: int = 0,
         cube: int = 0,
-        target_pe: int | str = 0,
+        target_pe: int | tuple[int, ...] | str = 0,
     ) -> Tensor:
         """Set DP placement metadata (like torch.Tensor.to())."""
         if placement is None: