GH-31: fix index race between concurrent instances

[blainemotsinger]

Fixes #31.

A second knowledge-mcp instance sharing the single-writer index blocked in bleve. Open (bbolt flock has no timeout) before serving MCP, so clients hit their handshake timeout.

• search.NewLazyIndex + OpenBackground: the index opens on a background goroutine after the MCP server starts; single-flight, self-recovering
• WaitOpen bounds tool waits: Query/Add return ErrIndexLocked (2s) instead of blocking forever while another instance holds the lock
• the pending open completes on its own once the lock frees (no restart)

Tests: 3 new in search/race_test.go (locked open, query while locked, add while locked). Full suite, go vet, gofmt, and -race pass. Manual e2e (two
live servers, shared index): locked handshake 0.02s, bounded error at 2.0s, self-recovery after the first instance exits.