chore(frontend): bump vllm rust crates to 60857baa

[FeathBow]

Description

Closes #1007

• Bump the five pinned vLLM Rust crates from d3e2888c (2026-08-25) to 60857baa (vllm main as of 2026-09-02).
• Picks up [Perf][Rust Frontend] Count the tokenizer vocabulary once at construction vllm-project/vllm#54449: the tokenizer wrapper counts its vocabulary once at construction instead of cloning the merged vocabulary on every request, which is what frontend: the tokenizer wrapper counts its vocabulary on every request #1007 measured on the serving path.
• No bridge changes: the EngineCore impersonation compiles unchanged against the new revision. The tokenizer crate's tekken migration drops tiktoken-rs, rubato, ndarray and hound from the lock.
• Mac pegainfer-sim A/B at 128-in / 64-out / 3200 prompts / c=320, alternated twice: 0-fail both sides; req/s 372.8 / 370.4 (d3e2888c) vs 374.9 / 374.7 (this branch); TPOT p50 12.16 vs 12.15 ms, the sim floor; TTFT p50 8.9 / 8.7 vs 8.5 / 8.6 ms.
• Tiny-prompt A/B on the 26B-A4B checkpoint (<bos>hi, one token, 5 warm + 30 timed, median end-to-end, main vs branch alternated twice): 50.5 / 52.4 ms on d3e2888c vs 11.3 / 11.5 ms on this branch, the per-request vocabulary clone frontend: the tokenizer wrapper counts its vocabulary on every request #1007 measured (48.2 ms → 12.5 ms with the cached count applied by hand) now gone through the pin.

Dependency Update

Crate	Previous revision	New revision
vllm-chat	d3e2888c	60857baa
vllm-engine-core-client	d3e2888c	60857baa
vllm-server	d3e2888c	60857baa
vllm-text	d3e2888c	60857baa
vllm-tokenizer	d3e2888c	60857baa

Notable upstream in the range (417 commits): the tokenizer vocabulary count cached at construction (#54449), the LogprobsTensors wire schema fix (#53939), the SSE streaming hot-path optimization (#51321), truncate_prompt_tokens / truncation_side support (#48584), stop_token_ids validated against the vocabulary (#54196), gRPC LoRA lifecycle control (#52840), the migration to the new tekken crate (#53056), decoded text attributed to tokens (#52910), bounded recursive argument parsers (#54303), and the DeepSeek V4 renderer fixes (#53281, #54854).

Test plan

• cargo test --release --locked -p pegainfer-frontend --lib — 67 passed
• cargo test --release --locked -p pegainfer-sim --lib — 6 passed
• cargo test --release --locked -p pegainfer-sim --test frontend_e2e — 13 passed
• Mac sim vllm-bench A/B vs the d3e2888c binary, same flags as above — 3200/0 both, throughput and TPOT flat (numbers above); cargo fmt --check and clippy --all-targets -D warnings clean for pegainfer-frontend and pegainfer-sim
• Single GPU (sm_89, x86_64), Gemma 4 26B-A4B NVFP4: cargo test --release --locked -p pegainfer-frontend --lib — 67 passed; tiny-prompt A/B above — main 50.5 / 52.4 ms, branch 11.3 / 11.5 ms (p90 within 1.3 ms of the medians on every run)