vulkan: parallelize DSV4 Lightning Indexer prefill

[Mushoz]

Same disclaimer as my previous PR :)

I have measured the following performance improvements:

209 -> 214 improvement at 32k
175 -> 190 improvement at 64k

As this PR improves the lightning indexer, the improvement will grow as context depth increases as the time spent on the lightning indexer increases with increasing context length. Micro benches confirm that the performance increase hold at least until 512k context length:

Source depth	kv	Baseline	Optimized	Reduction
0	512	4.23 ms	2.14 ms	49.5%
8k	2560	17.10 ms	10.09 ms	41.0%
16k	4608	29.97 ms	17.62 ms	41.2%
32k	8704	50.51 ms	32.94 ms	34.8%
64k	16896	94.36 ms	64.87 ms	31.3%
128k	33280	174.96 ms	128.76 ms	26.4%
256k	66048	352.73 ms	250.88 ms	28.9%
512k	131584	696.79 ms	495.90 ms	28.8%

It's a single commit this time.

[Mushoz]

Whoops, accidentally posted while I was still writing the description. Will edit the PR. My apologies for the early post.

Edit: Done!