fix(lsp): don't flag imports as unused when imported file defines a service or error

[klaidliadon]

Summary

Aggregator schemas like main.ridl that only compose other .ridl files were flagged "is unused" for every import, because the check only counted struct/enum/typealias references. Services and top-level errors are always-used contributors to the aggregated webrpc output, so their mere presence in the imported file means the import is live.

The same check also incorrectly suggested narrowing imports to a member list when the imported file contained a service or error — doing so would silently drop them from codegen.

Changes

• importContributionsOf (new helper): separates what an imported file contributes into three buckets — types (referenceable per name), errors (always-used), services (always-used).
• importDiagnostics now short-circuits when the imported file contributes a service or error, and only runs per-name reference counting on pure type-provider files.
• Adds three focused tests covering the aggregator / service / error cases.

Decision matrix after this change

Imported file defines…	Referenced by importer	Result
Any service	—	no warning
Any top-level error	—	no warning
Only types, none referenced	0	"is unused"
Only types, some referenced	partial	"can be narrowed to: …"
Only types, all referenced	full	no warning
Nothing	—	no warning

Selective imports (import "x.ridl" - Foo) are untouched — the transitive re-import check continues to work as before.

Real-world motivation

On omsx's schema/main.ridl — a pure aggregator that imports 11 service-bearing .ridl files plus errors.ridl — every import was flagged "unused" in the IDE. After this PR only a genuinely dead import (a model struct with no references) remains flagged, which is the signal the diagnostic is meant to produce.

Test plan

• New unit tests cover aggregator + service / errors-only / service-with-types-not-all-used
• Existing unused-import and narrowing tests still pass
• make test clean
• No new make lint findings (11 pre-existing failures in internal/ridl/parser.go and internal/lsp/semantic_document.go are untouched)