llama.cpp fixes variable-expert loading and backend paths before 0.4.0
Five India-date-pure llama.cpp builds refactored SYCL code, added optional fused residual operations, fixed loading for models with per-layer expert counts such as Nemotron Puzzle, corrected an s390x quantization accumulator and made build-info output configurable.
Why it matters: Loader correctness and backend-specific kernels can determine whether open mixture-of-experts models start, produce valid output or regress on less common hardware without any change to model weights.