[Tools][Qwen4Exp] Add QSA corpus and quality helpers

[Leonccaa]

Summary

This stacked follow-up keeps optional Qwen4Exp QSA calibration-corpus and
quality-evaluation helpers separate from the runtime-focused 1Cat PR.

It adds:

• a disjoint long-context corpus builder with overlap gating;
• a deterministic tool-call calibration corpus builder;
• a frozen-token runner for matched KV-cache quality comparisons;
• focused tests and usage documentation for those helpers.

Relationship to the core PR

• Base: feat/qwen4exp-qsa-e4m3-kv-core
• Upstream draft: 1CatAI/1Cat-vLLM#447
• This PR is not required to build, load, or run the calibrated E4M3 QSA
  implementation in [Model][SM70][FP8] Add calibrated E4M3 QSA KV cache 1CatAI/1Cat-vLLM#447.
• The core PR retains qsa_kv_calibration.py, including overlay generation.
• This follow-up contains no checkpoint overlay, calibration dataset, raw call
  log extractor, or environment-specific material.

Validation

• ruff-check: passed
• ruff-format: passed
• typos: passed
• check-spdx-header: passed after applying the required headers
• Focused tests: 8 passed in 4.88s
• Public-content scan: no internal hostnames, private addresses, user paths,
  Codex-specific references, or credential markers found

AI assistance

OpenAI Codex assisted with implementation review, validation, and PR
preparation. The submitter reviewed the resulting diff and test evidence.