Add reasoning-pattern benchmark scaffold

[LIU-ZHIYUAN-source]

Summary

This PR adds an initial benchmark scaffold for comparing inferlet-based reasoning patterns in Pie.

It includes:

• A reasoning benchmark inferlet with Direct, Best-of-N, Tree-of-Thought, and Graph-of-Thought modes
• A GSM8K-style benchmark runner with deterministic answer extraction/evaluation
• A CLI execution mode that runs through pie run with local wasm + manifest + config
• Strict final-answer parsing to avoid extracting numbers from incomplete generations
• A no-thinking control path for Qwen3-style models, currently implemented as a benchmark-local workaround

Validation

Tested locally / on RunPod with:

• cargo fmt --manifest-path inferlets/reasoning-benchmark/Cargo.toml
• cargo test --manifest-path inferlets/reasoning-benchmark/Cargo.toml
• python3 -m unittest benches.test_reasoning_bench
• RunPod smoke tests through ./target/release/pie run

Small sanity experiments:

• Qwen3-0.6B and Qwen3-1.7B on GSM8K-10
• Patterns: Direct, Best-of-N, Tree-of-Thought, Graph-of-Thought
• No runner errors observed
• Qwen3 no-thinking mode suppressed <think> after the benchmark-local cue workaround

Notes

This PR focuses on benchmark infrastructure, not final benchmark results. The current GSM8K-10 runs are only sanity checks and are not statistically meaningful.

Follow-up

This initial scaffold keeps all reasoning patterns in one benchmark inferlet for rapid validation. A planned follow-up is to split the patterns into separate inferlets and introduce a base text-completion inferlet, so user-submitted test-time scaling inferlets can be run and scored in isolation.