fix(pglite-sync): don't filter move-in messages by LSN

[jbingen]

Summary

Move-in messages from Electric's tagged_subqueries feature don't have an lsn header because they come from direct database queries, not from the PostgreSQL replication stream. Previously, these messages were incorrectly skipped as "already seen" because the missing LSN defaulted to 0.

This fix checks for the is_move_in header and bypasses LSN filtering for move-in messages, ensuring that rows moving into a shape due to subquery condition changes are properly synced to the client.

Problem

When using shapes with subqueries (e.g., WHERE id IN (SELECT user_id FROM workspace_members WHERE workspace_id = $1)), new rows that "move in" to match the subquery weren't being synced:

1. Electric server correctly processes the move-in and sends an INSERT message
2. The INSERT message has headers.is_move_in: true but NO headers.lsn (since it's from a query, not replication)
3. pglite-sync defaults missing LSN to 0
4. The check lsn <= lastCommittedLsnForShape evaluates to 0 <= X which is true after initial sync
5. The move-in message is skipped as "already seen"

Solution

const isMoveIn = (message.headers as Record<string, unknown>).is_move_in === true

if (!isMoveIn && lsn <= lastCommittedLsnForShape) {
  // Skip only if NOT a move-in message
  return
}

Test case

1. Create a shape with a subquery WHERE clause (e.g., profiles where id IN (SELECT user_id FROM workspace_members WHERE workspace_id = ...))
2. Complete initial sync
3. Add a new row to the inner table that causes a row to "move in" to the outer shape
4. Before fix: the moved-in row is not synced to PGlite
5. After fix: the moved-in row is correctly synced

[jbingen]

Added a second commit to handle duplicate key errors that can occur when move-in data overlaps with initial sync data.

When a row "moves in" to a shape, the server sends an INSERT. But if that row was already synced during initial sync, we get a duplicate key error.

For move-in inserts specifically, use ON CONFLICT DO UPDATE to upsert instead of plain INSERT. This only applies when is_move_in: true - regular inserts still use plain INSERT to catch any unexpected duplicates as errors.