Fix runtime wrong-result failure in test_swap.py

[xlinsist]

Deliverables

• Submit a PR on fixing the runtime wrong-result bug in the Triton RISC-V swap example.
• Include focused pytest evidence showing that python/examples/test_swap.py::test passes and that both swapped tensors match the expected originals.

Background

The current repository has a runtime correctness failure in test_swap.py. The first assertion confirms that x receives the original y, but the second assertion fails because y does not receive the original x:

assert torch.equal(y, x_)
E       assert False

Unlike the compiler failures, this case reaches execution and produces an incorrect result, so it should be treated as a generated-code or memory-aliasing correctness bug.

Task Description

• Reproduce from the triton-riscv repository root:

cd python/examples
pytest -q test_swap.py::test

• Inspect test_swap.py and the generated IR/object behavior for the swap kernel to determine whether the failure is caused by load/store ordering, aliasing, vectorization, or backend memory lowering.
• Implement the minimal fix that preserves both values during the swap and does not rely on changing the test expectation.
• Run the focused pytest case above and one nearby load/store or vector add example to check for shared memory-lowering regressions.
• In the PR description, include before/after pytest output and the root cause of the incorrect y value.

[xbwcan]

The reason is that the input memref was directly converted using bufferization.to_tensor, without using memref.alloc + memref.copy to preserve the original value.