Fix: Add BF16 conversion support for oneAPI SYCL backend

[yingying0906]

When running FLUX2-dev on Intel GPU with oneAPI SYCL backend, I got this error:

stable-diffusion.cpp/ggml/src/ggml-sycl/ggml-sycl.cpp:2047: GGML_ASSERT(to_fp32_sycl != nullptr) failed

The issue is that ggml_get_to_fp32_sycl() and related functions were missing BF16 conversion cases. Added the missing BF16 support in ggml/src/ggml-sycl/convert.cpp (Intel oneAPI only, guarded by __INTEL_LLVM_COMPILER).

Tested with flux2-dev-Q2_K.gguf and flux2-dev-Q4_0.gguf, both fp16 and fp32 build options - all working now.

Note: The actual code changes are in the ggml submodule.
View the changes here: yingying0906/ggml@75b092ba
Or see the branch: https://github.com/yingying0906/ggml/tree/fix-sycl-bf16

[Green-Sky]

You should upstream them to llama.cpp.

[yingying0906]

Thanks for the feedback! I'll submit this to llama.cpp as suggested.