[SYCL free func] migrate AmpKernels kernels

[minmingzhu]

AmpKernels.cpp — Kernel UT Report

• SYCL kernel conversion check: In src/ATen/native/xpu/sycl/AmpKernels.cpp, amp_update_scale_ff_kernel (line 118) is implemented/launched as a free function via SYCL_EXT_ONEAPI_FUNCTION_PROPERTY((syclexp::nd_range_kernel<1>)) and sycl_kernel_submit<amp_update_scale_ff_kernel>(...) function pointer.

• The remaining *Functor structs in this file (AmpNonFiniteCheckUnscaleFunctor, AmpForeachNonFiniteCheckUnscaleFunctor) are element-wise operation functors passed to gpu_kernel() / multi_tensor_apply() (TensorIterator/foreach infrastructure) and are not SYCL kernel launch functors.

• Changed-kernel related UTs run in both locations:

1. In PyTorch (/workspace1/minminz/workpath/pytorch-public):
   • pytest test/test_xpu.py -v -k "grad_scal or GradScal or amp"
   • Result: 16 passed, 170 deselected, 2 xfailed.
2. In current repo (torch-xpu-ops):
   • pytest test/xpu/test_autocast_xpu.py -v
   • Result: 26 passed, 3 skipped.
3. Functional verification (GradScaler step/update on XPU):
   • Normal path: scale remains 4.0 ✓
   • Inf-detected path: scale correctly backed off to 2.0 ✓