Extend cross-source finding dedup to custom analyzer sentences

[mamu0]

Extends the finding deduplication introduced for Advisor/AZQR/custom sources to also cover duplicate sentences produced by the custom (live-scan) analyzers themselves.

Previously, every sentence emitted by the per-resource analyzers (disk.ts, vm.ts, container-app.ts, etc.) shared the generic code: "custom.unknown" and had no recommendationId, so two custom findings describing the same problem on the same resource could never be collapsed — the existing dedup logic deliberately skips findings without a
recommendationId to avoid merging unrelated sentences on a guess.

Sentences that match a known template now get a stable, per-sentence custom.<id> identity (e.g. custom.disk.unattached, custom.low-cpu-usage), following the same <source>.<id> convention
already used by advisor.<id> and azqr.<id>. This lets the existing (already source-agnostic) fold logic collapse duplicate custom findings too, not just cross-source ones. Sentences with no known template are unaffected: they keep code: "custom.unknown" and no recommendationId, and are still never collapsed on guesswork.

Depends on #2026

Resolves: CES-2227