refactor: introduce AgentCliGenerator base to unify CLI generator handling

[omkargaikwad23]

Summary

• Add AgentCliGenerator base class (ABC) that the four CLI generators
  (gemini_cli, claude_code, codex_cli, agy_cli) now inherit from, declaring
  the shared interface: version, create_command, safe_generate,
  parse_response, extract_tools, extract_skills.
• Unify create_command to one signature across all four generators so the
  evaluator no longer branches on concrete type.
• Simplify AgentEvaluator: collapse the four isinstance(...4-tuple...)
  checks to a single isinstance(self.generator, AgentCliGenerator), replace
  the per-type version chain with self.generator.version, and drop all
  concrete CLI imports.

Why

Each new CLI generator previously required edits at ~6 sites in the evaluator.
Keying off the base class makes adding a generator zero-touch in the evaluator,
and the ABC enforces the interface at instantiation.

Test plan

• pytest test/agy_cli_test.py test/gemini_cli_test.py — 52 passed
• All four subclasses report no remaining abstract methods
• agentevaluator.py imports cleanly, no concrete CLI references
• Dockerised run logs: https://paste.googleplex.com/4915247398912000

[prernakakkar-google]

/gcbrun

[prernakakkar-google]

Please manually test against docker and attach results here.

[omkargaikwad23]

/gcbrun

[omkargaikwad23]

/gcbrun

[omkargaikwad23]

Please manually test against docker and attach results here.

Added. Please check.