OpenCL: allow mixed f16/f32 add

[rmatif]

Enables mixed-precision F16/F32 addition and fixes the use of LoRAs on sdcpp leejet/stable-diffusion.cpp#757

[lhez]

Apologize for the delay.

The mixed f16/f32 path is for dst type of f16, does not affect dst type of f32. Did some verification on A830 with language models - all looks good. There might be slow down since there are branches in the kernels (should be uniform, but may still affect the compiler). We can further iterate if needed.

[rmatif]

Apologize for the delay.

The mixed f16/f32 path is for dst type of f16, does not affect dst type of f32. Did some verification on A830 with language models - all looks good. There might be slow down since there are branches in the kernels (should be uniform, but may still affect the compiler). We can further iterate if needed.

I thought of the branching, but since it's not a critical op, I think it's fine. I didn't want to duplicate the code and add a new kernel just for this