refactor(board): simplify app and controller internals

[dgageot]

The pkg/board package had several repeated patterns that made the code harder to follow and introduced subtle maintenance hazards. Column lookup was duplicated across four methods using identical slices.IndexFunc calls, stored-project access was guarded by the same if a.config.Board != nil block in four places, and the per-card controller state was split across three parallel maps under the same mutex, which could drift out of sync.

This PR introduces two small private helpers in app.go — columnIndexLocked and projectsLocked — to centralize those repeated patterns, and consolidates the three per-card maps in controller.go (expectTurn, launchFailures, launchErrors) into a single map[string]*cardState. The forget() method now drops one entry instead of three, eliminating the map-drift risk entirely.

All changes are strictly behavior-preserving refactors with no functional modifications.

[docker-agent]

Assessment: 🟢 APPROVE

This is a clean, behavior-preserving refactor. The two new helpers (columnIndexLocked and projectsLocked) correctly centralize repeated patterns without altering semantics. The consolidation of three per-card maps into a single map[string]*cardState simplifies forget() and eliminates map-drift risk without introducing any correctness issues.

Key equivalences verified:

• AddProject reorder: when Board == nil, projectsLocked() returns nil; the duplicate loop over nil is a no-op — same outcome as iterating an empty Projects slice.
• setExpectTurn(false): now creates a *cardState entry instead of deleting; turnExpected correctly returns false via the s != nil && s.expectTurn check.
• resetLaunchFailures: setting to 0 vs deleting is functionally identical since callers only care whether the count is non-zero.
• setLaunchError(nil): entry persists with a nil error; LaunchError() correctly returns nil either way.