feat(inference): 引入串行/行并行/树并行推理后端并按后端拆分 CI

[huangbogeng]

Summary

• 增加编译期可选且互斥的推理后端特性：
  • infer-serial（默认）
  • infer-row-parallel
  • infer-tree-parallel
• 将推理内核重构为后端分发模式，XgbModel::predict_dense 公共 API 保持不变。
• 实现行级并行（微批优先，含小批量回退串行阈值）与树级并行（稳定优先：固定分块 + 确定性归并 + 补偿求和）。
• 增加树并行重复运行确定性测试，确保稳定输出。
• CI 从 --all-features 改为按后端矩阵分别执行 clippy/test，并同步更新 README、CONTRIBUTING、PR 模板和 CHANGELOG。

Motivation

面向低延迟/微批推理场景，需要在吞吐与稳定性之间提供可控选择。
通过编译期特性切换后端，可避免运行时策略分支开销，并确保部署形态可预测。

Validation

• cargo fmt --all
• cargo clippy --all-targets -- -D warnings
• cargo test --all-targets
• cargo clippy --all-targets --no-default-features --features infer-row-parallel -- -D warnings
• cargo test --all-targets --no-default-features --features infer-row-parallel
• cargo clippy --all-targets --no-default-features --features infer-tree-parallel -- -D warnings
• cargo test --all-targets --no-default-features --features infer-tree-parallel

Notes

• predict_dense 对外接口未变，更改集中在内部后端实现与构建特性。
• 后端特性被显式约束为“仅可启用一个”，若配置冲突会在编译期报错。
• 由于后端特性互斥，CI 不再使用 --all-features。

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c3a6008de6

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Decouple tree chunking from Rayon thread count

predict_dense_scalar_tree_parallel derives chunk_count from rayon::current_num_threads(), so the chunk boundaries (and therefore floating-point summation order) change when the pool size changes, e.g. different RAYON_NUM_THREADS values or custom pool sizes across environments. Because IEEE-754 addition is non-associative, the same model/input can produce different predictions across deployments, which conflicts with the deterministic backend goal and can flip multi:softmax decisions near ties. Use a chunking strategy that is stable for a given model (or make the thread-count dependency explicit and tested) to keep outputs reproducible.

Useful? React with 👍 / 👎.