Upgrade github.com/llm-d/llm-d-kv-cache-manager import to v0.3.0

[vMaroon]

Summary

This PR upgrades the github.com/llm-d/llm-d-kv-cache-manager import the the v0.3.0 release.

Note that this does not utilize the new preprocessing package for chat-completions support - this will followup in v0.4.0.

Benchmarking & Tests

• Overall v0.3.0 includes significant test coverage boosts
• Benchmarking data on the 37-capacity setup (open link for context)

Summary across QPS

Experiment	Output toks/s	Requests/s	Success Rate	TTFT p90 (s)	TTFT mean (s)	ITL mean (s)	ITL p50/ p90 (s)
precise v0.3.0-rc1	5533.7	6.920	100.00%	0.222	0.169	0.019	0.0000/0.094
precise v0.2.1	5650.0	6.914	100.00%	0.275	0.193	0.020	0.0000/0.085

EPP Queue and KV Cache Metrics Summary

Experiment	Wait Queue (mean/p90/max)	KV Cache % (mean/p90/max)	Pods	Data Points
precise v0.2.1	0.2/0/11	49.8/70.5/76.9	4	20824
precise v0.3.0-rc1	0.0/0/1	49.0/67.0/83.9	4	52920

Cool graphs

[nirrozenbaum]

/lgtm
/approve