[SYCL] Fix kernel name mangling for CUDA/HIP offload with shared libraries

SYCL kernel names are generated during template instantiation using
MangleContext. When using SYCL offload to CUDA/HIP, kernel names must
be consistent across compilation units, including shared libraries
compiled separately.

Previous approach only handled cross-ABI scenarios (Microsoft host +
Itanium device). This left same-ABI scenarios (Itanium host + Itanium
device) broken for shared libraries, as each compilation unit would
generate different lambda numbering, resulting in runtime errors:

  No kernel named _ZTSZZ21performIncrementationENK... was found

Solution:
For SYCL offload compilation, always use createDeviceMangleContext()
instead of createMangleContext(). Device mangling context uses
DeviceLambdaManglingNumbers which provides globally consistent lambda
numbering across compilation units.

This ensures identical kernel names for:
- Cross-ABI: Microsoft host + Itanium device (Windows + CUDA/HIP)
- Same-ABI: Itanium host + Itanium device (Linux + CUDA/HIP)
- Shared libraries compiled separately (.so/.dll)

Fixes CMPLRLLVM-69642
Fixes CMPLRLLVM-69415

Author: kekaczma