ci: build and test Paddle custom operator with CUDA

The Paddle border-op CPU-branch regression (source/tests/pd/test_border_op.py) cannot fail in any current CI pipeline:

- `.github/workflows/test_cuda.yml` sets `DP_ENABLE_PADDLE: "0"` at workflow level, so both GPU jobs build without Paddle.
- `.github/workflows/test_python.yml` installs Paddle from the **cpu** nightly index.
- `copy_local_tensor_data` in source/op/pd/comm.cc is compiled only under `GOOGLE_CUDA`/`TENSORFLOW_USE_ROCM`.

Net effect: no configuration builds Paddle with CUDA, so the operator path touched by #5832 is neither compiled nor executed in CI. A regression to an unconditional `gpuMemcpy` would go undetected.

Proposal: add a CI configuration (or a manual/GPU workflow) that builds the Paddle custom operator with CUDA so the `test_border_op_self_copy_uses_cpu_place` CPU-branch test and the rest of source/op/pd/ get real coverage.

Related: #5832