llama.cpp: update submodule for CPU fallback fix

[cebtenzzre]

There was a small error in #2409 that caused the Kompute build of llama.cpp to attempt to allocate tensors on GPU, even when we intended to fall back to CPU due to missing features.

This fixes the crash when attempting to load a model that is not supported by Kompute:

ggml_backend_sched_backend_from_buffer: error: no backend supports buffer type Kompute1 used in tensor blk.0.attn_q.weight
GGML_ASSERT: /home/jared/src/forks/gpt4all/gpt4all-backend/llama.cpp-mainline/ggml-backend.c:1115: false