Fallback fft_r2c

[CuiYifeng]

• Removed conditional compilation logic for USE_ONEMKL_XPU in _fft_r2c_xpu and _fft_r2c_xpu_out functions to perform the FFT computation on the CPU and transferring the result back to the XPU.
• Added a test case, test_mem_eff_attention_fail_with_batch_size_geq_65536, to the skip list due to a known issue with CUDA not being enabled.

[Copilot]

Pull Request Overview

This PR refactors the XPU FFT routines to always use a CPU-based MKL fallback for _fft_r2c_xpu and _fft_r2c_xpu_out, removing the USE_ONEMKL_XPU conditional paths and consolidating logic for CPU computation followed by device transfer.

• Always call native::_fft_r2c_mkl on CPU and transfer result to XPU.
• Removed conditional headers and unified includes for FFT ops.
• Simplified output resizing and copying logic in the _out variant.

Comments suppressed due to low confidence (3)

src/ATen/native/xpu/SpectralOps.cpp:6

• These FFT includes (_fft_c2c_native.h and _fft_c2r_native.h) are not used in the r2c fallback logic and can be removed to reduce unused dependencies.

#include <ATen/ops/_fft_c2c_native.h>

src/ATen/native/xpu/SpectralOps.cpp:90

• Consider adding or updating tests to cover this CPU-to-XPU fallback path to ensure correctness and performance characteristics on XPU when MKL is not available.

Tensor out_cpu = native::_fft_r2c_mkl(

src/ATen/native/xpu/SpectralOps.cpp:87

• [nitpick] The CPU fallback transfer logic is duplicated across both functions; consider extracting the pattern (CPU compute + XPU copy) into a shared helper to reduce duplication.

Tensor _fft_r2c_xpu(