[executorch][native-ggml] Size the scheduler hash for graph leafs too

## Ulterior Motive

Running Gemma 4 31B through ExecuTorch Native's GGML engine on CUDA, to find why it decodes slower than llama.cpp on the same GGUF.

## Rationale

**What**: The backend scheduler is created for `2 * graph_size` tensors.

**Why**: `ggml_backend_sched` hashes every node and every leaf of a graph it splits, and asserts that its hash set holds them all. It was sized for `graph_size`, which caps nodes and leafs separately, so a graph whose nodes and leafs together outgrew it failed that assertion, though either count alone fit.

## Details

- **Before**: `ggml_backend_sched_new(..., graph_size, ...)`; ggml's own source notes the hash "needs to be size*2 to account for leafs".
- **After**: `2 * graph_size`, room for a full graph of `graph_size` nodes and `graph_size` leafs.

`test_graph_with_as_many_leafs_as_nodes` adds 2,100 separate constants to an input: 2,100 nodes and as many leafs, within the default graph size of 4,096.

Authored with Claude Code.

Differential Revision: [D130721991](https://our.internmc.facebook.com/intern/diff/D130721991/)


ghstack-source-id: 469829014
Pull-Request: #24847

Author: SS-JIA

backends/native/runtime/engine/ggml/GgmlExecutor.cpp

@@ -492,11 +492,13 @@ class GgmlExecutable final : public EngineExecutable {
     std::ranges::transform(backends_, raw.begin(), [](const auto& backend) {
       return backend.get();
     });
+    // The scheduler hashes a graph's leafs as well as its nodes, and a graph
+    // holds up to graph_size of each.
     ggml_backend_sched_ptr scheduler(ggml_backend_sched_new(
         raw.data(),
         nullptr,
         static_cast<int>(raw.size()),
-        options_.graph_size,
+        2 * options_.graph_size,
         /*parallel=*/false,
         /*op_offload=*/true));
     if (!scheduler) {

backends/native/test/runtime/ggml/test_ggml_ops.py

@@ -1510,6 +1510,18 @@ def test_index_put_reads_transposed_values_without_copies(self):
         )
         self.assertEqual(ops["CONT"], 0)
 
+    def test_graph_with_as_many_leafs_as_nodes(self):
+        # 2,100 additions of separate constants fit the default graph size of
+        # 4,096 nodes, but the scheduler also hashes their 2,100 leafs.
+        constants = [torch.randn(4) for _ in range(2100)]
+
+        def forward(x):
+            for constant in constants:
+                x = x + constant
+            return x
+
+        self._check(Lambda(forward), (torch.randn(4),))
+
     def test_cached_graphs_refresh_inputs_and_persistent_cache(self):
         t = Dim("t", min=1, max=8)
         ep = torch.export.export(