fix: avoid RequestManager deadlock in LSP rename by using doSyncRead

[mvanhorn]

Summary

LSP rename, prepareRename, and quickfix change conversion no longer queue a nested read request and block on it, which removes the RequestManager deadlock window: a nested read [R2] waits on a queued write [W], which awaits the in-flight outer read [R1], which waits on [R2]. The three call sites now use the existing ILanguageServerAccess#doSyncRead, which performs the same read without queueing.

Why this matters

#3735 traces both in-tree paths precisely: RenameService2.rename() and prepareRename() call doRead(uri, ...).get() from inside the read request that LanguageServerImpl already queued, and ChangeConverter2 does the same inside the code-action request. The reporter's company hits this regularly through slow quickfixes. The issue body itself proposes doSyncRead as the fix.

Changes

• RenameService2.rename() restructures the .exceptionally(...).get() chain into a try/catch that preserves the existing contract: null on a root-cause FileNotFoundException when shouldPrepareRename is true, everything else rethrown via Exceptions.sneakyThrow.
• ChangeConverter2 switches its doRead(...) + immediate join to doSyncRead. Every in-tree caller (the quickfix path through DiagnosticResolution and QuickFixCodeActionService) already runs inside a request and the method blocked on the future immediately, so the synchronous read is behavior-preserving apart from removing the deadlock window, answering the question the issue raised about whether the async form was ever required.

Testing

New regression test in Rename2Test exercises rename while requests interleave; the existing rename, prepareRename, and code-action tests in org.eclipse.xtext.ide.tests stay green.

Fixes #3735

[github-actions]

Test Results

  8 055 files  + 1    8 055 suites  +1   4h 6m 13s ⏱️ + 18m 0s
 43 168 tests + 1   42 577 ✅  -  6    584 💤 ±0  7 ❌ +7 
212 750 runs   - 18  209 825 ✅  - 22  2 918 💤  - 3  7 ❌ +7 

For more details on these failures, see this check.

Results for commit ed81a47. ± Comparison against base commit 66bc577.

♻️ This comment has been updated with latest results.

[mvanhorn]

Pushed ed81a47 fixing the test failures: the deadlock regression test bound InterleavingRequestManager inside Rename2Test, which overrode the IRequestManager binding for every test in that class and broke the sibling rename tests. The harness and the regression test now live in a dedicated Rename2ConcurrencyTest with its own injector, so Rename2Test runs against the standard RequestManager again. Also working on getting the ECA registered for my commit email.