Expose opt-out for horizontal partition fusion (pytorch#23082)

Summary:

`CapabilityBasedPartitioner` normally merges supported partitions even when their nodes are independent in dataflow. This produces larger delegated subgraphs, but its pairwise candidate checks are quadratic in the number of partitions. Wide graphs with many independent branches—Eagle-3 is the motivating case—can therefore spend prohibitive time in partition generation before backend compilation begins.

`CapabilityBasedPartitioner` already exposes `skip_horizontal_fusion`, but the canonical `generate_pattern_op_partitions` helper did not forward it. Backend integrations had to duplicate the helper or accept the quadratic work. Expose the option in both the fbcode and xplat mirrors; the default remains false, so existing callers keep their current partitioning behavior.

Differential Revision: D121425768

Pull Request resolved: pytorch#23082

Author: ThomasJannaud

exir/backend/canonical_partitioners/pattern_op_partitioner.py

@@ -110,6 +110,7 @@ def generate_pattern_op_partitions(
     partitions_list: Optional[List[List[torch.fx.Node]]] = None,
     op_support: Optional[OperatorSupportBase] = None,
     ignore_literals: bool = False,
+    skip_horizontal_fusion: bool = False,
 ) -> List[Partition]:
     """
     Args:
@@ -126,6 +127,7 @@ def generate_pattern_op_partitions(
             - Getting the result of create_op_support()
             - Getting the result of create_pattern_support()
             - Multiple OperatorSupportBase classes chained together with chain()
+        skip_horizontal_fusion: Restrict merging to data-dependent partitions.
 
     Returns
         A list of partitions (largest possible subgraphs) containing nodes are
@@ -175,6 +177,7 @@ def is_node_supported(self, submodules, node: torch.fx.Node) -> bool:
         graph_module,
         final_op_support,
         allows_single_node_partition=True,
+        skip_horizontal_fusion=skip_horizontal_fusion,
     )
     partition_list = capability_partitioner.propose_partitions()