Dequant improvements rebase

[AidanBeltonS]

This PR provides improvements to the dequantize_block_q4_K kernel. It focuses on improving the global memory accesses.

Three main changes are implemented:

• Single 32 bit load for half2 rather than two 16 bit loads
• Load all scales in to local memory then do random access on results
• Vectorize the q load so we load 32bits each time rather than 8bits

All results below collected on A100 GPU

		Without Changes	With Changes	% Change
LLama-bench 70 B	PP Throughput (t/s)	503.36	564.04	-11.85	Negative change is better
	NSYS Avg Kernel time (us)	587.54	409.52	30.30	Positive change is better

No meaningful change in Intel GPU results have been observed.

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

[OuadiElfarouki]

Improvement observed on Nvidia A4000 & RTX 4070 as well (7B & 13B - Q4_K_*).
Thanks!

[joeatodd]

Ping @airMeng to check for regressions on Intel side

[airMeng]

It is weird that I don't see any performance improvements on Arc A770, no regression either.

ping our performance expert @luoyu-intel

[luoyu-intel]

The code looks good. It can prevent cache misses. So you may not see the performance improvement if there are no cache misses in your case.