gooddata-eval: drop the dead `skills` fallback spelling in set_skills parsing

[Tomkess]

Split out of #1762 discussion (agreed there as separable).

_set_skills_declarations() in packages/gooddata-eval/src/gooddata_eval/core/agentic/conversation.py hedges across two argument spellings:

names = args.get("skill_names")
if names is None:
    names = args.get("skills")

skill_names is the real key: the set_skills tool declares skill_names: list[str] and its run() reads that field. Nothing emits or reads a bare skills. Verified against the gen-ai service while answering the replace-vs-append question on #1762.

So the fallback is provably dead. It was also actively harmful once: test_agentic_conversation.py's _skills_tc helper emitted {"skills": [...]}, and every conversation test passed through the fallback, exercising a payload the platform never sends. Fixed in #1762 (9965722e), but the fallback is what let it go unnoticed — a hedge that silently accepts a wrong shape hides test bugs rather than tolerating real traffic.

Scope:

• Remove the or args.get("skills") branch.
• Grep for remaining "skills"-as-argument-key spellings in tests. test_visualization_evaluator.py and test_agentic_visualization.py already use skill_names and are correct.

Worth doing after (or together with) the result-vs-arguments change, since that touches the same parse.