Refactor/converge cold build

[JesseHerrick]

dexter init and the LSP server both build an index from nothing, and each had its own implementation. init parses on every core and writes one bulk transaction with the indexes dropped; the server walked serially, parsed one file at a time, and committed a transaction per file against live indexes. On a 10k-file corpus the server's path took 4.441s against the shared pipeline's 1.088s, for identical rows.

Move init's pipeline into internal/indexer and call it from both. The server reaches it only under IsEmpty(), which is where its own build already lived.

Three things do not transfer from init --force, and this is why the change is a shared pipeline rather than a shared function call:

• Deleting the database files and reopening. The server holds the store and handlers read that pointer unsynchronised. It is also pointless: an empty index has nothing stale to throw away.
• SetBulkPragmas. Leaving WAL needs exclusive access, so journal_mode fails whenever another connection is open. Options.InProcess suppresses all four. They were the smallest part of the win; the parse pool and the dropped indexes are the rest of it.
• Insert-only mode's assumptions. It skips the per-file DELETE and allocates file ids from a counter seeded when the batch opens, so a save landing mid-build would duplicate rows or collide on a primary key. The server now holds indexWrites for writing across a full build, and every single-file write — save, watched file, rename — holds it for reading. Incremental sweeps are unaffected: they use the per-file path and take no lock.

Two fixes fall out of it:

• The prune step deletes every stored path the walk did not see. A full build populates no seen set, so running the prune after one would delete the index that was just written. It is now inside the incremental branch only.
• SetBulkPragmas applied synchronous before journal_mode, so a locked database failed after disabling fsync and left it disabled for the life of the process. Harmless in a CLI that exits; not harmless in a server. journal_mode goes first, so a failure changes nothing.

The LSP path also records IndexVersion now, which only init used to do.