[codex] Load plugins and skill roots concurrently

[mzeng-openai]

Why

Configured plugins and skill roots are independent reads, but the cold startup path loads them serially. Bounded concurrency reduces that latency without allowing one large catalog to create unbounded filesystem work.

What changed

• load up to eight configured plugins concurrently
• scan up to eight skill roots concurrently through the shared PluginSkillRootCache
• use order-preserving buffering so configured-plugin order, root priority, deduplication, and warning resolution remain deterministic
• box the bounded worker futures so downstream spawned session futures remain compiler-tractable
• add coverage that configured plugin results remain sorted

This is stacked on #28623, which introduces the shared plugin skill-root cache used by this change.

Validation

• just test -p codex-core-skills -p codex-core-plugins (359 passed)
• just fix -p codex-core-skills -p codex-core-plugins
• just bazel-lock-check