lhez

Properly identify mark multi rope and vision rope and mark them as unsupported so that these rope variants get put back to CPU and does not crash. Also `fp16` variant...

Currently small models like qwen2.5 0.5B does not work properly with OpenCL backend. This PR fixes this issue. This PR also changes subgroup size to 64 for all Adreno GPUs.

The rms_norm_mul kernel produces incorrect result when ne00 = 768. This PR changes how the kernel does reduction to get the sum. This seems to fix the issue.