Faster CAS hits during builds

CAS (compilation cache) operations now skip the per-op fsync on the local analytics database, removing a stall that made each materialize hit very slow during builds. Previously, every CAS op enqueued a SQLite insert that fsynced, tying up the daemon’s cooperative thread pool (and getting worse the slower the underlying disk was at fsyncing, e.g. hosted runner VMs). The analytics DB has no durability requirement, so it now opens with synchronous = OFF. Effect: per-op CAS latency drops substantially across every build using the compilation cache, not only on hosted macOS runners but on any self-hosted setup where disk fsync contended with compiler output during a build.