refactor(workspace): reduce unused public surface

[xiaguan]

Summary

• reduce frontend, model, KV, and kernel visibility using all-features Hawk reachability
• delete dead state, wrappers, and test probes exposed by the visibility reduction
• retain pegainfer-bench because Qwen3 and Kimi-K2 report binaries are real consumers
• fix the Hawk playbook so the Triton Python path remains valid from build-script working directories

Hawk

Finding	main	This PR	Delta
dead_public	27	26	-1
unnecessary_public	462	321	-141
unnecessary_restricted_visibility	87	14	-73
Total	576	361	-215

Both sides were measured from the same main commit. Of the 361 remaining findings, 332 belong to the deliberately excluded kvbm-logical fork. pegainfer-bench has zero findings.

Validation

• cargo fmt --all --check
• cargo check --release --workspace --all-features --all-targets
• CPU CI Clippy package gate with --all-targets -- -D warnings
• Qwen/CUDA CI Clippy package gate with --all-targets -- -D warnings
• cargo test --release --workspace --lib
• final all-features cargo hawk check

The broader all-features workspace Clippy command still reaches pre-existing pedantic errors in unchanged GLM kernel and test code; both actual CI Clippy package gates pass.