Retain rejected output in TypeScript and Python failures

[robgruen]

Summary

Fixes #116.

• Add optional data: unknown and responseText to TypeScript failures, and value: object, response_text, and has_value to Python failures.
• Retain rejected inputs in the TypeScript, Zod, program, and Python validators, including when those validators are called directly.
• Preserve custom failure metadata and the latest useful rejected candidate when a repair request fails or returns unparseable output.
• Keep successful result shapes, human-readable diagnostics, exception behavior, and the existing language-specific repair policies unchanged. Retain bounded failure context, not an attempt history.
• Document payload availability, success discriminators, cross-attempt context, and sensitive-output handling in package READMEs and public usage guides.

Validation

• TypeScript source build and test compilation passed.
• Focused TypeScript failure, TypeScript validator, Zod, program, and model regression suites passed.
• Python failure, translator, validator, and model tests: 44 passed; 5 existing snapshots unchanged.
• Pyright over the Python source and affected callers: 0 errors or warnings.
• git diff --check passed.

Compatibility

Failure payloads are unvalidated and must not be treated as the success type. Use success in TypeScript and isinstance in Python to distinguish outcomes. Python's has_value distinguishes unavailable data from explicit None or falsy values. The latest response text may be from a later repair attempt than the retained rejected candidate.

Matching .NET implementation: microsoft/typechat.net#338 (standalone PR against main).