refactor(qwen3): move the report harness out of the library API

[onenewcode]

Description

Fixes #944.

pegainfer-qwen3 exposed pub mod kernel_bench (1112 lines) as public API. Every normal library and server build compiled report-only benchmark code and shipped its types in the crate's public surface.

The harness moves to pegainfer-qwen3/src/bin/report_support/, split across five modules, and is #[path]-included by the two kernel-report binaries. It is now compiled per binary and never as part of the library.

• qwen3_kernel_report takes the whole tree. qwen3_model_report needs only common::L2CacheClear and includes that module alone, with the #[allow(dead_code)] on its own mod declaration rather than inside the shared module — so the other four modules stay dead-code-checked by the kernel report's build. Removing the allow and building qwen3_kernel_report alone yields 0 dead-code warnings, confirming the allow covers only the model report's narrower use.
• The model report imports production primitives from their owning crates: build_split_kv_csr from pegainfer_core::ops, split_qkv_into from pegainfer_kernels::ops, instead of reaching through a harness re-export.
• SplitKvConfig::{new, label} widen to pub. parse was already public and calls new, so a private constructor only pushed external callers onto the unchecked struct-literal path; label is parse's inverse, and restating its format in the harness would silently break the round-trip contract that label_parse_round_trips guards.
• SPLIT_KV_CHUNK_TOKENS / SPLIT_KV_TUNED_MAX_CHUNKS are re-exported through runtime so the harness keeps deriving the Tuned split width from the runtime constants exactly as it did before the move, rather than restating 64x64 where a later retune would silently desync the report's split_tmp_* sizing.

Report schemas, CLI flags, benchmark cases, numeric policy, and kernel selection are unchanged; both binaries' --help output is identical.

Incidental fix. qwen3_model_report did not compile under --features kernel-report at main: it called ops::split_qkv_into, which pegainfer-core does not re-export. This was invisible because no CI job enables kernel-report — the two binaries carry required-features = ["kernel-report"], and qwen3-cuda-clippy does not pass it (the feature pulls pegainfer-cupti, excluded at ci.yml:300 for missing CUPTI headers in the CI toolkit). Worth a follow-up: this PR does not create that gap, but it does move ~1500 more lines behind it.

Docs. docs/models/qwen3/roadmap.md cleanup ledger strikes the kernel_bench.rs file-size entry. docs/subsystems/kernels/kernel-op-reports.md TL;DR records the new home and notes that paths in its Execution Log are historical — those are dated past-tense step records (Step 1 still uses the pre-flattening crates/ layout), so rewriting them to today's paths would falsify the record rather than fix it.

Type of Change

• Bug fix (non-breaking change which fixes an issue) — the qwen3_model_report compile break described above
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected) — the crate's public API loses pub mod kernel_bench. No runtime behavior changes and no in-workspace consumer breaks; flagged because Move the Qwen3 benchmark harness out of the library public API #944 mandates the removal with no compatibility shim, so any out-of-tree user of that path is affected.
• Documentation update

Checklist

• My code follows the style guidelines of this project (see docs/conventions/coding-style.md).
• I have performed a self-review of my own code.
• I have formatted my commits according to Commitizen conventions.
• I have run the local test suite and all tests pass (see CLAUDE.md).

Run locally on RTX 5090 (sm_120), driver 610.57.04:

Gate	Result
cargo check --release -p pegainfer-qwen3 --lib	pass, 0 warnings
cargo check --release -p pegainfer-qwen3 --all-targets --features kernel-report	pass
cargo clippy --release -p pegainfer-qwen3 --all-targets --features kernel-report -- -D warnings	pass
cargo test --release -p pegainfer-qwen3 --lib	88 passed, 0 failed
cargo build --release -p pegainfer-qwen3 --features kernel-report --bin qwen3_kernel_report --bin qwen3_model_report	both link
cargo fmt --check	pass