sycl: refactor quantization to q8_1

[Alcpz]

The current implementation of how some mul_mats do 8-bit quantization is not very flexible. While exploring other possibilities for a different gemv kernel, I run into the necessity of having a q8_1 tensor in a slightly different format, and that wasn't supported with the current convert_src1_to_q8_1 bool.

The PR refactors quantization kernels to a separate header and:

• Unifies kernel submission to use sycl::nd_item<1>
• Rewrites the quantize_q8_1 to have the same structure as the reorder q8_1 kernel
• Adds exception handling that was ignored in the original code introduced with SYCLomatic.

Performance is unaffected.

Pinging @AD2605 as author of the reorder q8_1 kernel.