sycl-exp : Re-enabled mul_mat_batched_sycl Path for batched Q*K & KQ*V

[OuadiElfarouki]

This PR re-enables the MUL_MAT path for QK and QKV operations involving MKL's gemm_batch through ggml_sycl_mul_mat_batched_sycl(otherwise unused as the src1->type() == GGML_TYPE_F16 is never realized) .
Prompt Processing performance is restored, for e.g. :

• Nvidia A100 + 70B Q4_K (b=p=512): 241 t/s -> 538 t/s
• Intel Arc A770 + 13B Q4_K (b=p=512) : 363 t/s -> 587 t/s
• Nvidia A4000 + 7B Q4_K (b=p=512) : 1710 t/s -> 2033 t/s

Text Generation performance remained the same / showed very small improvements.

• I have read the contributing guidelines
• Self-reported review complexity:
  • Low
  • Medium
  • High

[joeatodd]

@airMeng this PR might restore the perf regression you describe here?

[airMeng]

Outstanding to catch this small mistake!

[joeatodd]

Outstanding to catch this small mistake!

Thanks for the feedback @airMeng - be aware this PR is to our temporary perf branch codeplay/sycl-main. You will want it in master too I guess!