Add model-agnostic analysis, response-cleaning, prompt/version tracking, incremental runs, and Chinese prompts

[Huberyky]

Motivation

• Enable non-OpenAI (OpenAI‑compatible) models by restoring cost estimates and sanitizing vendor reasoning traces so downstream JSON parsing is robust.
• Provide built‑in reliability/validation/robustness analytics so multi‑run outputs can produce standard measurement diagnostics used in social‑science research.
• Improve reproducibility and incremental workflows by binding prompt versions to outputs, reporting parse success rates, and supporting prompt language selection (e.g., Chinese templates).

Description

• Added a new analysis module gabriel.analysis exposing reliability, validate, and robustness helpers for Krippendorff's α, ICC, Pearson/Spearman, MAE, F1/Cohen's κ, stratified reports, and bootstrap CIs.
• Introduced gabriel.utils.model_utils with strip_reasoning_tags, prompt_hash, write_run_metadata, and load_incremental_cache to strip vendor CoT traces (e.g., <think>...</think>), compute prompt hashes, and manage incremental reuse.
• Extended core tasks (rate, classify, extract) to accept prompt_language and incremental flags, write prompt‑hash metadata to run_metadata.json, sanitize raw responses with strip_reasoning_tags before parsing, produce per‑task parse reports (*_parse_report.csv), and attempt incremental merged reads when incremental=True.
• Added Chinese Jinja2 prompt templates (ratings_prompt_zh.jinja2, classification_prompt_zh.jinja2, extraction_prompt_zh.jinja2) and wired template selection by prompt_language.
• Added pricing entries for DeepSeek / Qwen families to MODEL_PRICING so cost estimates remain available when using OpenAI‑compatible base_urls.
• Small plumbing: exported new top‑level functions (gabriel.reliability, gabriel.validate, gabriel.robustness) and added unit tests and utilities accordingly.

Testing

• Ran targeted unit tests: pytest -q tests/test_analysis_extensions.py tests/test_imports.py tests/test_discover_exports.py, all tests passed.
• Performed static sanity checks: python -m py_compile $(rg --files src/gabriel -g '*.py' | tr '\n' ' ') succeeded (no syntax errors).
• Verified template and runtime integration by running the new analysis and response‑sanitization flows in task code paths (parse reports created and incremental merge logic exercised during tests).

---

Codex Task