Harden CybORG agent validation

[Muhtasham]

Summary

• instantiate submitted CybORG MyAgent during validation using the same constructor fallbacks as the runtime
• reject submissions that import and subclass correctly but would crash immediately during episode setup
• add regression coverage for runtime-style construction and constructor failures

Why

A real OpenAI smoke run generated a CybORG agent that passed validation but crashed during evaluation because its constructor called RandomAgent.__init__(seed=...). The arena isolated the crash and scored it correctly, but validation should catch this before the round starts.

Verification

• uv run pytest -q tests/arenas/test_cyborg.py tests/arenas/test_scml.py -> 14 passed
• uv run ruff check codeclash/arenas/cyborg/cyborg.py tests/arenas/test_cyborg.py -> passed
• uv run ruff check . -> passed
• uv run pytest -q -> 187 passed
• uv run python main.py configs/examples/CybORG__dummy__r1__s2.yaml -o /tmp/codeclash-cyborg-dummy.pUV7lx -> passed full Docker smoke, per-episode results had status: ok
• uv run python main.py configs/examples/SCML__dummy__r1__s2.yaml -o /tmp/codeclash-scml-dummy.WT6xqD -> passed full Docker smoke, per-sim score details written
• long local SCML smoke: tournament.rounds=5, sims_per_round=5, n_steps=8, output /tmp/codeclash-scml-long.CZnbSS -> passed 6 competition executions with 30 total SCML worlds
• long local CybORG smoke: tournament.rounds=5, episodes_per_round=5, steps_per_episode=10, output /tmp/codeclash-cyborg-long.ZgGyM7 -> passed 6 competition executions with 30 total CybORG episodes, all visible episode records had status: ok

Note: I also ran short real OpenAI-backed smoke loops for SCML and CybORG to verify the coding-agent path. Those are integration smoke evidence only, not benchmark results.

CI note: pre-commit and test pass on this PR. markdown-link-check currently fails on an unrelated pre-existing 404 in codeclash/viewer/_STATIC_README.md.