Keep LSP workspace preparation off the protocol loop

Follow-up opportunity

Nonblocking architecture/performance follow-up from #505, verified at 429c883a. The prepared-workspace handoff is a useful boundary, but expensive preparation still runs synchronously on the protocol thread. This is confirmed by the call path; no large-workspace latency benchmark was run for this review.

Evidence

• iris-lsp::start creates a current-thread Tokio runtime.
• initialized and finish_workspace_configuration call apply_configuration_inner through synchronous router notification/event handlers in compiler-bin/iris-lsp/src/server.rs.
• apply_configuration_inner performs discovery, source reconfiguration preparation, and initial build_initial before returning. discover_manual also waits for std::process::Command::output() synchronously.
• build_initial / query_package waits for Rayon package execution; query_package at lines 330–341 drives queries through JavaScript generation. Parallel package work does not make the calling protocol handler asynchronous.
• Introduced by package-scheduled LSP bootstrap, retained after workspace runtime encapsulation. The test requests_are_cancelled_while_the_workspace_is_loading checks a waiting State directly, not protocol responsiveness while preparation is executing.

Why it matters

While a preparation handler is running, the main loop cannot service incoming requests, shutdown, or newer configuration messages. The existing waiting-state rejection and notification queue help while waiting for client configuration, but cannot run concurrently with synchronous preparation. Slow source commands can extend this indefinitely.

Bounded completion criteria

• Move discovery/initial compilation and source-reconfiguration preparation into owned background work; return results through a generation-tagged event and commit only a current result on the protocol thread.
• Preserve ordered pending notifications, open buffers, previous-workspace retention on preparation failure, and the existing distinction between preparation and delivery failure.
• Add a protocol E2E test with deterministically blocked preparation that proves the loop can respond before preparation is released, and that a superseded result cannot replace the newest configuration. Define shutdown cleanup for outstanding preparation/source-command work.

Duplicate check: searched open and closed issues for initialization, blocking, responsiveness, and configuration. Closed #47 concerns cancellation versus file locking, and closed #497 establishes workspace settings; neither covers asynchronous preparation.

Start with compiler-bin/iris-lsp/src/lib.rs and server.rs, especially initialized, finish_workspace_configuration, and apply_configuration_inner; then inspect build_initial and query_package in compiler-bin/iris-build/src/compile.rs. Run the existing requests_are_cancelled_while_the_workspace_is_loading test before adding a protocol E2E test with blocked preparation. Done means the protocol loop remains responsive, superseded results are ignored, existing notification and workspace behavior is preserved, and shutdown cleans up outstanding work.