Propagate fatal errors from async/cont callbacks

[arimu1]

Description

Fixes #4505

Fatal errors (OutOfMemoryError and other non-UnsafeNonFatal throwables) delivered through an async/cont resume callback were completed as Outcome.Errored on the fiber instead of crashing the process. That differed from IO.raiseError(fatal) and from fatals thrown inside IO.delay / IO thunks, which already call onFatalFailure.

This shows up with IO.fromCompletableFuture because CompletableFuture.handle catches the fatal and resumes with Left(t), and also with a minimized form:

IO.async_[Unit] { cb => cb(Left(new OutOfMemoryError("Boom!"))) }

Approach

Detect fatal errors as soon as the cont/async result is consumed by the runloop, matching the existing Error / raiseError case:

• asyncContinueFailedR — async resume with Left(fatal)
• IOCont.Get when the callback already completed — synchronous cont completion with Left(fatal)

No change to fromCompletableFuture itself: rethrowing inside CompletableFuture.handle is unnecessary once the fiber treats the resumed fatal correctly, and avoids leaving the cont suspended if the throw is mishandled on the CF thread.

Prior attempt #4512 targeted a similar issue but was closed (hanging tests / abandoned). This PR only touches the fiber paths that consume the failure.

Testing

• Added IOApp process tests (recommended for fatals so the shared test runtime is not shut down):
  • exit on fatal error from async_
  • exit on fatal error from CompletableFuture (JVM)
• Local: sbt ioAppTestsJVM/testOnly cats.effect.IOAppSpec — 28 examples, 0 failure (Temurin 21), including both new cases

[selormwalker]

Nice improvements! Ready for merging.