fix(resolve): memoize an import's resolved target (#636)

[someshSandbox]

Fixes #636.

A view with many exposes and a metadata usage stalls validation:

metadata def Tag;
package P1 { }
package P2 { }
// … P12
view v {
    expose P1::*;
    expose P2::*;
    // … expose P12::*;
    @Tag;
}

exposes	before	after
10	3 s	0.03 s
11	31 s	0.03 s
12	no result in 60 s	0.03 s
200	—	0.04 s

Cause: each expose's target was resolved again on every lookup in the view, never memoized (it resolves under inAllVisible), so every lookup searched the other exposes in every order.

Change: a new importTargets table keeps each import's resolved target.

• Only hits are kept. A miss can still mean a sibling import was suspended.
• It is never read or written under InCondition (review finding).
• It is journaled like implicitParams and counted in MemoSize.

This matches the OMG pilot, where an import's target is a linked cross-reference resolved once.

Tests:

• tests/model/expose_scaling_test.go: 20 exposes must analyse within 10 s. Fails on develop, passes here (0.5 s).
• filter_test.go, TestAnImportTargetFoundWhileAConditionIsResolvedIsNotRemembered: a target found in condition mode doesn't leak to an ordinary lookup.
• internal/semantic/..., tests/model/... and internal/workspace/... all pass on develop with this change.
• gofmt is clean. make lint wasn't run (tools not installed).

Does not fix #633 (a public import cycle, a different path).

🤖 Generated with Claude Code