Expose --queue-size, --queue-timeout-secs, and --rate-limit-tokens-per-second CLI flags

[pedramr]

Purpose

The standalone Rust vllm-router binary hardcodes three queue / rate-limiting
knobs in CliArgs::to_router_config() and exposes no CLI flags for them:

• queue_size = 100
• queue_timeout_secs = 60
• rate_limit_tokens_per_second = None

…even though RouterConfig already supports all three fields and the Python
launcher (py_src/vllm_router/router_args.py) already exposes them as
--queue-size / --queue-timeout-secs / --rate-limit-tokens-per-second.
This is a flag drift between the two entry points: configuration that works via
the Python launcher silently cannot be set when running the binary directly.

In particular, binary users cannot set --queue-size 0, which disables the
concurrency queue so the limiter sheds immediately with HTTP 429 (fail-fast)
instead of queuing up to 100 requests for up to 60s.

This PR adds the three flags to CliArgs and threads them through
to_router_config(). Defaults are identical to the previously hardcoded
values (100 / 60 / None), so behavior is unchanged unless a flag is
explicitly passed. The new --help text mirrors the Python launcher's wording.

Test Plan

Built and exercised the release binary with stable Rust (rustc 1.96.0):

• cargo fmt --check
• cargo check
• cargo clippy --all-targets
• cargo build --release
• ./target/release/vllm-router --help lists the three new flags with defaults
• Boot the binary with --queue-size 0 and confirm RouterConfig::validate()
  accepts it

Test Result

• cargo fmt --check, cargo check, cargo clippy --all-targets, and
  cargo build --release: all clean.
• --help output:
  • --queue-size <QUEUE_SIZE> — [default: 100]
  • --queue-timeout-secs <QUEUE_TIMEOUT_SECS> — [default: 60]
  • --rate-limit-tokens-per-second <RATE_LIMIT_TOKENS_PER_SECOND> — falls back
    to --max-concurrent-requests when unset
• Booting with --queue-size 0 logs Configuration validated successfully and
  proceeds past config validation (the limiter then runs queue-less, shedding
  with HTTP 429 at the concurrency limit).

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR — fix the binary/Python-launcher flag drift so the
  queue and rate-limit knobs are configurable from the binary.
• The test plan — cargo fmt/check/clippy/build --release plus a --help
  and --queue-size 0 boot check.
• The test results — all checks clean; --help and --queue-size 0 boot
  verified (see above).