FEAT: add ANSI escape output scorers

[Utkarsh Bahuguna (u7k4rs6)]

Summary

Adds dedicated output scorers for detecting ANSI escape sequences in model outputs, based on the detection patterns used by Garak's AnsiEscape detector.

Changes

• Added AnsiEscapeOutputScorer

  • Detects raw ANSI escape sequences and C1 control characters in output.
  • Covers CSI (ESC[) and OSC (ESC]) sequences, as well as C1 CSI/OSC characters.

• Added EscapedAnsiOutputScorer

  • Detects ANSI escape sequences represented as escaped/literal text.
  • Covers \x1b, \033, \u001b, \27, \e, \x9b, and \x9d representations.

• Added scorer exports through the existing PyRIT scoring package.

• Added Garak provenance metadata for the detector patterns.

• Added comprehensive unit tests covering:

  • Positive and negative cases
  • Garak payload parity
  • Raw vs. escaped payload separation
  • Adversarial/non-matching inputs
  • Custom patterns
  • Scoring rationale
  • Memory behavior
  • Python string escaping edge cases

• Updated the scorer documentation and corresponding Jupyter notebook.

Scope

This PR implements the scorer portion of the ANSI escape support described in #2683. The Garak AnsiEscape scenario and techniques are intentionally left for a follow-up PR.

Part 1 of #2683.

[Roman Lutz (romanlutz)]

Is this a duplicate of #2688 ?

[Utkarsh Bahuguna (u7k4rs6)]

Roman Lutz (@romanlutz) No, they’re separate. #2688 adds the ANSI output scorers, while #2689 addresses a different part of the ANSI-related work. There’s no duplicate implementation between the two.

[Roman Lutz (romanlutz)]

Identical PR title and the same classes mentioned in the description suggest otherwise (?)

Please make the distinction clear.

[Utkarsh Bahuguna (u7k4rs6)]

Right, there's significant overlap.
One way I could separate them is to keep AnsiEscapeOutputScorer in #2688 and make #2689 specifically the EscapedAnsiOutputScorer + its tests/integration. Would that distinction make sense, or would you prefer me to close #2689?