[codex] Add SA-Bench DeepSeek V4 tokenizer mode

[alec-flowers]

Summary

Adds a narrow SA-Bench tokenizer-mode pass-through for DeepSeek V4 without bringing in any recipe payload.

Current main already has the DSV4 custom tokenizer package and the fast chat-template failure with guidance from PR #76. This PR fills the remaining gap from aflowers/gb200-dsv4-recipes: recipes can now set benchmark.tokenizer_mode: deepseek_v4, and SA-Bench will pass that through to benchmark_serving.py.

What changed

• Adds BenchmarkConfig.tokenizer_mode.
• Passes tokenizer mode from SABenchRunner into bench.sh as a new trailing positional argument, preserving the existing dataset arguments.
• Adds --tokenizer-mode deepseek_v4 support in benchmark_serving.py and backend_request_func.py.
• Adds custom_tokenizer: deepseek_v4 as a short alias for the same vLLM DeepSeek V4 tokenizer.
• Adds focused tests for command construction.

Validation

• PYTHONPATH=src .venv/bin/python -m pytest tests/ -q -> 612 passed, 2 skipped, 6 deselected
• PYTHONPATH=src .venv/bin/python -m pytest tests/test_benchmarks.py tests/test_configs.py -q -> 104 passed
• UV_PROJECT_ENVIRONMENT=.venv uv run --frozen --no-sync ruff check src/srtctl tests/test_benchmarks.py --ignore SIM117 -> passed
• UV_PROJECT_ENVIRONMENT=.venv uv run --frozen --no-sync ruff format --check src/srtctl tests/test_benchmarks.py -> passed
• bash -n src/srtctl/benchmarks/scripts/sa-bench/bench.sh -> passed
• PYTHONPATH=src .venv/bin/python -m py_compile src/srtctl/benchmarks/scripts/sa-bench/backend_request_func.py src/srtctl/benchmarks/scripts/sa-bench/benchmark_serving.py -> passed

Note: this intentionally does not port the side-branch auto-fallback behavior for missing chat templates; current main already fails fast with actionable DSV4 guidance instead.