trunk/d9ae411a24a7346c1140865b821b0a5e06b59c6e: Skip Quack RMSNorm override on ROCm and non-CUDA tensors (#185644)

Summary:
The Quack RMSNorm override in _native/ops/norm/rmsnorm_impl.py is selected for any tensor whose dtype is fp16/bf16/fp32 and whose torch.cuda.get_device_capability major is 9 or 10.

Two cases slip through that shouldn't:

• ROCm builds (e.g. MI300x). get_device_capability on gfx942 returns major=9, matching the H100 check, so the dispatcher routes to Quack - a CUTLASS/CUDA-only kernel that has no business running on AMD.
• CPU tensors. _is_supported never checks input.device.type. When a CPU tensor is passed (e.g. during the FX Interpreter.run in splitting/utils.py:remove_unexpected_type_cast), get_device_capability reports the current CUDA device's capability, the cond passes, and the kernel crashes with ValueError: Mismatched Tensor ... expected device_type=cuda.

Test Plan: This fixed 3 tests internally

Differential Revision: D106733305

Pull Request resolved: #185644
Approved by: https://github.com/desertfire