[Feature][Frontend] Expose effective attention block size for DCP

[JulienDarve]

Purpose

With decode context parallelism (DCP), a physical KV cache block of 16 tokens can represent 64 tokens across four ranks. Clients need that effective size to interpret full-attention cache events.

This PR exposes the value as CacheConfig.effective_attention_block_size, populated from the initialized cache managers. Python clients read it from their config, and the engine startup response carries it to Rust and gRPC Control.GetServerInfo.

Existing physical block-size fields remain unchanged. Without DCP, the effective size matches the full-attention physical block size. The new field is optional and unavailable when the engines cannot report a common full-attention block size.

Duplicate checks found no open PR exposing this DCP-adjusted value through config, startup metadata, and gRPC.

AI assistance: OpenAI Codex assisted with implementation and validation.

Test Plan

Python:

PYTHONPATH=. .venv/bin/python -m pytest \
  tests/config/test_config_utils.py \
  tests/v1/core/test_kv_cache_utils.py \
  tests/v1/engine/test_engine_core_client.py \
  -k 'effective_attention_block_size or apply_ready_response or cache_config_hash' \
  -q --tb=short

From rust/:

cargo nextest run -p vllm-engine-core-client -p vllm-server \
  -E 'test(python_msgpack_fixtures) | test(grpc::tests::control_)'
cargo clippy -p vllm-engine-core-client -p vllm-server --all-targets -- -D warnings

From the repository root:

.venv/bin/pre-commit run mypy-3.12 --all-files --hook-stage manual
cargo fmt --manifest-path rust/Cargo.toml --all -- --check

Test Result

• Python: 9 passed, covering DCP=1/4 cache-manager values, startup synchronization, missing fields, rank disagreement, and compilation hash stability.
• Rust: 7 passed, covering MessagePack compatibility and gRPC control behavior.
• GPU smoke: hmellor/tiny-random-LlamaForCausalLM on RTX 6000 Ada generated four tokens. With DCP=1, the Python config and actual startup payload both reported 16 tokens; physical block size remained 16. The run used precompiled extensions and native sampling.
• Multi-GPU DCP inference was not run; DCP=4 was covered at the cache-manager level.
• Clippy, pre-commit hooks including manual mypy checks, protobuf lint, and protobuf backward-compatibility checks against the PR base passed.
• Backward compatibility: the base Python decoder accepted the captured new startup payload; the updated decoder accepted the base payload with the new field unavailable.
• Full-repository Python 3.12 type checking and Rust formatting passed.
• Fork CI checks are skipped because the workflows run only for vllm-project; the validation results above were obtained locally.