Use downcasting instead for accessing worktree-specific state of the Files.

This will allow us to introduce a WorktreeId type and use that everywhere
for identifying worktrees. It also just removes some unnecessary coupling
between the language crate and the worktree.