refactor(skills): share plugin namespace resolver

[anp-oai]

why

#31348 centralizes root-level plugin namespace resolution for ordinary skill loading, but the environment loader still carried a parallel namespace map, probe fanout, and ancestor-selection loop. Keeping both implementations would make namespace precedence fixes easy to apply in only one path.

what

• reuse SkillNamespaceResolver from the environment loader
• retain the environment loader's filter that avoids probing plugin roots without discovered skills
• keep namespace resolution concurrent with skill file reads
• delete the duplicate namespace map construction and per-skill ancestor walk

validation

• just test -p codex-core-skills namespace (10 passed)