ci: Add benchstat-based performance regression detection

[gyliu513]

Problem

We have Go benchmarks (make bench) but they are not integrated into CI. There is currently no automated way to detect performance regressions introduced by a PR. Maintainers must run benchmarks manually or rely on real cluster testing to validate performance, which slows down the development cycle.

Current state

• make bench only covers pkg/preprocessing/chat_completions/ and pkg/tokenization/
• 12 Benchmark* functions exist across 4 files, but several are not included in make bench
• No benchstat, no performance comparison tooling, no CI benchmark job

Proposal

Add a new GitHub Actions workflow (.github/workflows/ci-bench.yaml) that compares benchmark results between the PR branch and main using benchstat:

1. Run go test -bench=. -benchmem -count=5 on the PR branch → new.txt
2. Checkout main, run the same command → old.txt
3. Run benchstat old.txt new.txt and post the comparison as a PR comment
4. (Optional, future) Fail CI if any benchmark regresses beyond a configurable threshold

Benchmarks to track

Package	Key benchmarks
pkg/preprocessing/chat_completions/	BenchmarkRenderChat, BenchmarkRender, BenchmarkGetOrCreateTokenizerKey
pkg/tokenization/	BenchmarkAsyncTokenizationStress, BenchmarkSyncTokenizationStress
pkg/kvevents/	BenchmarkZMQSubscriber_Throughput
pkg/kvcache/kvblock/	(see "New benchmarks" below)

All benchmarks should use -benchmem to track allocation counts and catch GC pressure regressions.

CI performance impact

• Runs as a separate workflow (ci-bench.yaml), fully parallel with existing ci-test and ci-lint does not slow them down
• Can be configured as a non-required check so it never blocks PR merges
• Can be scoped to only trigger on changes to pkg/**/*.go, go.mod, go.sum

@vMaroon @sagearc comments for this new ci workflow?