Restart a crashed language server per --restart never|always|N

The process is supervised: on exit after a successful initialize it is
spawned again with exponential backoff (500ms to 30s), open documents are
re-sent as didOpen before the new session is announced, and requests caught
by the crash wait for it and retry. A crash more than a minute after the
previous one starts a fresh streak; reaching the streak limit (default 3)
ends restarting for good, after which every request fails with an LspError
saying so. Restarts and give-ups are logged and forwarded to the MCP client.

Writes to the child's stdin go through a PassThrough so a broken pipe ends
the session instead of surfacing as an unhandled rejection.

test/fixtures/server.ts

@@ -2,7 +2,8 @@
  * A miniature language server used by the tests. Words are symbols: hover
  * describes the word under the cursor, definition is its first occurrence,
  * references and rename cover every occurrence, and each line containing
- * "TODO" yields a warning. `fixture/state` reports open/close counts.
+ * "TODO" yields a warning. `fixture/state` reports open/close counts and
+ * `fixture/crash` kills the process.
  */
 import {
   createConnection,
@@ -136,6 +137,7 @@ connection.onDocumentSymbol(({ textDocument }) => {
   );
 });
 
+connection.onNotification("fixture/crash", () => process.exit(1));
 connection.onRequest("fixture/state", () => ({
   open: documents.all().length,
   closes,