Fix Inverse kernel rank underflow before indexing trailing dims (microsoft#28400)

### Description

`Inverse::Compute` (CPU and CUDA) read `dims[num_dim - 2]` and
`dims[num_dim - 1]` before validating `num_dim >= 2`. Because `num_dim`
is `size_t`, scalar or 1D inputs silently underflow the subtraction and
produce out-of-bounds indices.

- **`contrib_ops/cpu/inverse.cc`** — add `ORT_RETURN_IF_NOT(num_dim >=
2, ...)` immediately after shape retrieval, before any dimension
indexing.
- **`contrib_ops/cuda/inverse.cc`** — same guard in `ComputeInternal`.
- **`test/contrib_ops/inverse_test.cc`** — add `scalar_input_fails` and
`one_dim_input_fails` test cases.

### Motivation and Context

Without the guard, passing a 0-D or 1-D tensor to the Inverse op causes
`size_t` underflow on `num_dim - 2`, reading garbage memory before any
validation occurs. The fix is a defense-in-depth kernel-level check that
mirrors the existing ONNX shape-inference rejection.

Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: tianleiwu <30328909+tianleiwu@users.noreply.github.com>
Co-authored-by: Tianlei Wu <tlwu@microsoft.com>

Author: Copilot

onnxruntime/contrib_ops/cpu/inverse.cc

@@ -70,6 +70,8 @@ Status Inverse::Compute(OpKernelContext* ctx) const {
   const auto num_dim = input_shape.NumDimensions();
   auto* output = ctx->Output(0, input_shape);
 
+  ORT_RETURN_IF_NOT(num_dim >= 2, "Input tensor rank must be >= 2, got: ", num_dim);
+
   int64_t num_batches = 1;
   const int64_t rows = input_shape.GetDims()[num_dim - 2];
   const int64_t cols = input_shape.GetDims()[num_dim - 1];

onnxruntime/contrib_ops/cuda/inverse.cc

@@ -142,6 +142,8 @@ Status Inverse::ComputeInternal(OpKernelContext* ctx) const {
   const auto num_dim = input_shape.NumDimensions();
   auto* output = ctx->Output(0, input_shape);
 
+  ORT_RETURN_IF_NOT(num_dim >= 2, "Input tensor rank must be >= 2, got: ", num_dim);
+
   size_t num_batches = 1;
   const size_t rows = static_cast<size_t>(input_shape.GetDims()[num_dim - 2]);
   const size_t cols = static_cast<size_t>(input_shape.GetDims()[num_dim - 1]);

onnxruntime/test/contrib_ops/inverse_test.cc

@@ -86,5 +86,20 @@ TEST(InverseContribOpTest, four_by_four_batches_float) {
   test.AddOutput<float>("Output", {3, 4, 4, 4}, output);
   test.Run();
 }
+
+TEST(InverseContribOpTest, scalar_input_fails) {
+  OpTester test("Inverse", 1, kMSDomain);
+  test.AddInput<float>("X", {}, {4.f});
+  test.AddOutput<float>("Y", {}, {0.25f});
+  test.Run(OpTester::ExpectResult::kExpectFailure, "rank must be >= 2");
+}
+
+TEST(InverseContribOpTest, one_dim_input_fails) {
+  OpTester test("Inverse", 1, kMSDomain);
+  test.AddInput<float>("X", {4}, {4.f, 7.f, 2.f, 6.f});
+  test.AddOutput<float>("Y", {4}, {0.f, 0.f, 0.f, 0.f});
+  test.Run(OpTester::ExpectResult::kExpectFailure, "rank must be >= 2");
+}
+
 }  // namespace test
 }  // namespace onnxruntime