Update ggml_sycl_op_mul_mat_vec_q

[AidanBeltonS]

This PR updates the unsupported quantized data types and refactors the code for ggml_sycl_op_mul_mat_vec_q.
SYCL does not currently have the intrinsics to support some quantized data types, this adds one missing quantized data type to the unsupported check, so tests won't be run.
This also refactors the code so there is a single templated mul_mat_vec_q_sycl_submitter rather than multiple duplicate functions which submit a different instantiated kernel. This makes the code less verbose and much smaller.

[AidanBeltonS]

@NeoZhangJianyu, @abhilash1910, @Alcpz, feedback would be appreciated

[Alcpz]

Minor comment on the refactor. Looks great.

[abhilash1910]

Thanks @AidanBeltonS , could you please rebase to latest master for CI build?
Also tagging @ggerganov & @airMeng for a look when available.

[abhilash1910]

@ggerganov @0cc4m I think the vulkan build CI is exiting abruptly - maybe issue is common for other requests. Could you help take a look ? Thanks

[ggerganov]

It's because we enabled the gguf example which links only the ggml library and does not link ggml-vulkan or ggml-rocm:

#5216 (comment)

We can easily disable the gguf example from the build, but I'm wondering if we need separate libs in the first place. Can we not link everything in ggml, similar to what CUDA, Metal, SYCL, etc. do

[airMeng]

LGTM

[abhilash1910]

@AidanBeltonS could you please rebase to latest master - should solve some build issues with vulkan ci.

[NeoZhangJianyu]

1. "unsupported quantized data types"
   What's the data types to be supported by this PR?

2. How to test with the supported data type?

[AidanBeltonS]

1. "unsupported quantized data types"
   What's the data types to be supported by this PR?

   2. How to test with the supported data type?

1. This PR does not change the supported and unsupported data types. If you look at the switch case, all the types that are there are supported. It simply explicitly sets an assert for the quantization types we do not support, and sets tests which use that data type to unsupported.

2. All the regular tests for this functionality are run. So any existing testing for the other quantization types will still be run as normal to check functionality.