gooddata-eval: read set_skills' result rather than its arguments to determine active skills

[Tomkess]

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

_set_skills_declarations() / _final_skill_declaration() in packages/gooddata-eval/src/gooddata_eval/core/agentic/conversation.py derive the active skill set from the set_skills tool call's arguments — i.e. what the agent asked for. The same is true of _check_visualization_skill_activated in core/evaluators/visualization.py.

That is not always what became active. The service:

• drops requested names it does not recognise, and
• pulls in skills declared as dependencies of the ones requested.

The tool's own result echoes back the authoritative post-replacement set, so reading the result would be strictly truer to what actually became active than re-deriving it from the request.

Consequence today: a fixture requesting a misspelled or retired skill name is credited with activating it, and a skill activated only as a dependency is not credited at all — so both skill_routing and full_skill_coverage can be wrong in either direction.

Left out of #1762 because it is a behaviour change to how activation is measured, independent of the replace-semantics fix that PR is about, and it needs its own fixtures.

Scope:

• Parse the set_skills tool result and prefer it over the arguments, falling back to arguments when the result is absent or unparseable (older captured traces, and any trace where the call errored).
• Same treatment in core/evaluators/visualization.py.
• Tests covering: unrecognised name requested, dependency pulled in implicitly, and result missing (fallback path).