fix(core): release per-run process listeners and task history results - #36974

…#36866)

## Current Behavior

Every `runDiscreteTasks` call registers SIGINT/SIGTERM/SIGHUP handlers
(plus the forked runner's message/exit handlers) that are never removed.
Each closes over its orchestrator, so a long-lived process like an Nx
Cloud agent retains every orchestrator it ever created: the task hasher
(with its native project-graph copy), the lifecycles, and every task
result with full terminal output. `TaskHistoryLifeCycle.pendingResults`
also keeps every `TaskResult` for the life of the process. A profiled
customer agent grew from 237MB to 5.5GB over 50 tasks.

## Expected Behavior

Process listeners are removed once a run completes, task history stops
retaining terminal output, and agent memory stays flat across assignment
batches.

## Backport notes

Three places where 22.7.x differs from the branch this came from:

- `setupSignalHandlers` still registers three separate `once` handlers
  with different bodies here, rather than the single `handleSignal`
  helper. Each is hoisted to a named const so `dispose()` can remove it;
  the bodies are unchanged. A `once` listener that never fires stays on
  the emitter, so the leak is the same.
- `ForkedProcessTaskRunner.cleanup()` is synchronous on this branch, so
  the `await` in `dispose()` is a formality. The comment says the
  children are signalled first rather than reaped first.
- `packages/nx` runs on jest here, not vitest, so the three specs use
  `jest.*`. `task-orchestrator.spec.ts` does not exist on 22.7.x and is
  added with only the tests this change introduces.

(cherry picked from commit 0ae9069)

Co-authored-by: FrozenPandaz <jasonjean1993@gmail.com>
Co-authored-by: nx-cloud[bot] <71083854+nx-cloud[bot]@users.noreply.github.com>