Don't use dl_tensor.byte_offset when exporting capsules. (#21153)

Pytorch ignores byte_offset at least in some circumstances, meaning we
end up with the wrong data getting output.

This solves the issue by baking the byte_offset directly into the
returned pointer.

See:
https://github.com/dmlc/dlpack/blob/7f393bbb86a0ddd71fde3e700fc2affa5cdce72d/include/dlpack/dlpack.h#L225

Signed-off-by: Andrew Woloszyn <[email protected]>

Author: AWoloszyn

runtime/bindings/python/hal.cc

@@ -739,9 +739,13 @@ py::object HalDevice::CreateDLPackCapsule(HalBufferView& buffer_view,
       "Cannot export device buffer");
   static_assert(sizeof(dl_tensor.data) >=
                 sizeof(external_buffer.handle.device_allocation.ptr));
+  // Set the data pointer to the offset, and the byte_offset to 0.
+  // This SHOULD not be required, but some backends (torch GPU for example),
+  // ignore the byte_offset entirely.
   dl_tensor.data =
-      reinterpret_cast<void*>(external_buffer.handle.device_allocation.ptr);
-  dl_tensor.byte_offset = offset;
+      reinterpret_cast<uint8_t*>(external_buffer.handle.device_allocation.ptr) +
+      offset;
+  dl_tensor.byte_offset = 0;
 
   // Create and return capsule.
   PyObject* capsule = PyCapsule_New(static_cast<DLManagedTensor*>(tensor.get()),