[Performance]: Remove per-condition serialization in ConditionEvaluator under -mt

### Problem

Under `-mt`, `ConditionEvaluator` holds the per-condition expression-pool lock while evaluating, serializing identical imported conditions across projects.

### Proposed work

Narrow the `Stack` lock to pop/push, cap retained pool depth, add a parallel benchmark, and validate with same-base OrchardCore `-mt` measurements. Prototype: `origin/perf/AR-May/fix-condition-evaluation-in-mt`.

An 8-worker stress test improved condition throughput by 36–75%; end-to-end build gain is not yet measured.

Related: #14020, #12151.

Start by locating ConditionEvaluator and its per-condition expression-pool Stack, then inspect how the lock surrounds evaluation under -mt. Add the proposed parallel benchmark and compare same-base OrchardCore -mt measurements; done means lock scope is narrowed, pool depth is capped, and condition throughput and end-to-end effects are measured.

Written by the indexing model from the issue text.