perf(model): serve integration plan from the request scope

[bpamiri]

Part of #3213.

The cached component-integration plan (#3236) was re-read from the application scope on every model/controller/mapper materialization — roughly six synchronized application-scope lookups per instance. The plan is stable within a request (the only invalidation is the #3457 rebuild-on-poison validation, which replaces both copies), so after its first use per request it is now served from a request-scope struct.

Measured on the #3213 benchmark rig (Lucee 7, 2,000 model().new(), instrumented):

• plan lookup: 123 ms → 2 ms
• total instantiation: ~1058 ms → ~830 ms (~20%)

Full breakdown of the remaining per-object cost is in the #3213 thread; the copy loop (~45%) and instance setup (~40%) are the next targets.

IntegrationPlanSpec now also pins the request-cache behavior (second call returns the same plan array without touching the application scope) and the rebuild test clears the request cache so it still exercises the poisoned application-scope path.

Verification

• Local Lucee 7: global 243 pass, model 991 pass, 0 fail.
• Local RustCFML full suite: no new failures vs baseline.
• Full local Lucee suite + CI to follow.