Use matrix for gradient. (#9508)

- Use the `linalg::Matrix` for storing gradients.
- New API for the custom objective.
- Custom objective for multi-class/multi-target is now required to return the correct shape.
- Custom objective for Python can accept arrays with any strides. (row-major, column-major)

src/data/array_interface.h

@@ -384,7 +384,7 @@ inline bool ArrayInterfaceHandler::IsCudaPtr(void const *) { return false; }
  *   numpy has the proper support even though it's in the __cuda_array_interface__
  *   protocol defined by numba.
  */
-template <int32_t D, bool allow_mask = (D == 1)>
+template <std::int32_t D, bool allow_mask = (D == 1)>
 class ArrayInterface {
   static_assert(D > 0, "Invalid dimension for array interface.");
 
@@ -588,7 +588,7 @@ class ArrayInterface {
 };
 
 template <std::int32_t D, typename Fn>
-void DispatchDType(ArrayInterface<D> const array, std::int32_t device, Fn fn) {
+void DispatchDType(ArrayInterface<D> const array, DeviceOrd device, Fn fn) {
   // Only used for cuDF at the moment.
   CHECK_EQ(array.valid.Capacity(), 0);
   auto dispatch = [&](auto t) {

src/data/data.cc

@@ -448,7 +448,7 @@ void CopyTensorInfoImpl(Context const& ctx, Json arr_interface, linalg::Tensor<T
   auto t_out = p_out->View(Context::kCpuId);
   CHECK(t_out.CContiguous());
   auto const shape = t_out.Shape();
-  DispatchDType(array, Context::kCpuId, [&](auto&& in) {
+  DispatchDType(array, DeviceOrd::CPU(), [&](auto&& in) {
     linalg::ElementWiseTransformHost(t_out, ctx.Threads(), [&](auto i, auto) {
       return std::apply(in, linalg::UnravelIndex<D>(i, shape));
     });