Port roi_align (CUDA) to stable ABI. - #9573

Ports roi_align CUDA/ROCm kernels to the stable ABI.

Notes

• Requires [Stable ABI] Move fastAtomicAdd to torch/headeronly. pytorch#191501.
• Determinism layer change: The CUDA backward's at::globalContext().alertNotDeterministic call can't be made from a stable kernel, so it moves to the Python backward as torch._prims_common.alert_not_deterministic, gated to grad_output.device.type == "cuda". The CPU backward remains unchanged .

Stable-ABI audit (torch-abi-audit)

Ran torch-abi-audit on a CUDA build. The counts don't move with this layer since the CUDA kernel's unstable references are all shared with the remaining legacy ops.

    Package: torchvision
      Torch ABI:   UNSTABLE
      CPython ABI: n/a
      Bundled libs: 3
      -- bundled libs --
        [UNSTABLE] [not-abi3] _C.so            (stable_shim=0,  unstable=103)
        [STABLE  ] [not-abi3] _C_stable.so     (stable_shim=70, unstable=0)
        [STABLE  ] [not-abi3] image_stable.so  (stable_shim=76, unstable=0)