Capture RL trajectories and logprobs

[Lixtt]

Summary

• add opt-in RL trajectory JSONL capture for agent execution, LLM turns, tools, compaction, and terminal outcomes
• expose OpenAI-compatible logprobs/top_logprobs through Rust, Python, and Node SDK options
• rebase the work on A3S-Lab/Code main and keep the recorded tool definitions aligned with the actual LLM request
• add compatibility with the currently published a3s-memory 0.1.1 so Code 4.3.0 builds from public sources

Validation

• cargo fmt
• cargo clippy --workspace --lib --bins -- -D warnings
• cargo test -p a3s-code-core rl_trajectory
• cargo test -p a3s-code-core openai
• python3 -m maturin build --manifest-path sdk/python/Cargo.toml --release --out target/wheels
• python3 -m pip install --user --force-reinstall target/wheels/a3s_code-4.3.0-cp310-cp310-manylinux_2_35_x86_64.whl
• python3 import smoke test
• cargo build --manifest-path sdk/node/Cargo.toml --release
• cargo test --manifest-path sdk/python/Cargo.toml session_options_map
• cargo test --manifest-path sdk/node/Cargo.toml session_options_maps