llama.cpp b10353 fixes silent CUDA/Metal roll bug for accurate tensor ops

llama.cpp, the wildly popular C++ library for running LLMs locally (123k stars on GitHub), has shipped release b10353 with an important correctness fix. The release addresses a bug in the ROLL tensor operation where CUDA and Metal backends would silently produce incorrect results for non-contiguous source tensors. The ggml_roll function only asserted that nb[0] equals the type size, leaving permuted tensors technically valid inputs. However, the CUDA and Metal kernel implementations indexed strictly by element count (ne) and never read the stride (nb) values, causing mismatched output without any warning.

The fix adds a contiguity requirement to both CUDA and Metal backends, matching the same guard already used for GGML_OP_ROPE. This ensures the scheduler now falls back to the CPU implementation when a non-contiguous source is passed, which correctly handles strides. The release also includes a new permuted test_roll case to catch this class of bug in the future. This is a subtle but significant fix for production users—silent numerical errors in tensor operations are the worst kind, as they corrupt model outputs without crashing. Anyone running llama.cpp on NVIDIA GPUs via CUDA or Apple Silicon via Metal should update to b10353 to avoid unpredictable results when their models or applications use non-contiguous memory layouts.