llama.cpp b9933 fixes OpenCL Q6_K garbled output and buffer overlap

The llama.cpp b9933 release addresses critical OpenCL issues that caused garbled output for Q6_K quantized weights when the dimension ne01 is not a multiple of 128. This bug affected models like granite-3.1-3b-a800m-instruct, where odd vocab sizes triggered incorrect memory handling on Adreno GPUs.

The fix routes dense matrix multiplication away from the noshuffle path when ne01 % 128 != 0, using flat GEMV for decoding and falling back to CPU for other shapes. Additionally, the release implements get_alloc_size for OpenCL buffers to reserve alignment slack, preventing subbuffer overlaps that silently corrupted neighboring tensors.

This update ensures correct inference results on devices with specific alignment constraints and prevents memory corruption in quantized weight handling.