[release-1.13] reland the scheduler task and backport the interrupt fixes and scheduler optimization

[IanButterworth]

Since Keno's cancellation overhaul (#60281 and follow-ups) is not going to be backported to 1.13, this PR instead brings 1.13's interrupt handling to the quite healthy state (with new interrupt tests passing robustly) that master reached just before that overhaul landed.

Claude:

---

Reland (revert of #61065)

Reverts the four commits of #61065, restoring the scheduler-task work that was reverted on this branch while Ctrl-C delivery was broken (the breakage is fixed by the backports below):

• Scheduler: Use a "scheduler" task for thread sleep (Scheduler: Use a "scheduler" task for thread sleep #57544)
• simplify wait() (simplify wait() #58595)
• make sched_task more robust (make sched_task more robust #59669)
• the accompanying test compensation commit

Backports

In master order:

• Fix interrupting. Add tests. #62069 — Fix interrupting. Add tests. (fixes CTRL+C is even even worse on 1.14 #58689, the reason Scheduler: Use a "scheduler" task for thread sleep #57544 was reverted here)
• scheduler: avoid O(nthreads) wake-storm on every @spawn #61826 — scheduler: avoid O(nthreads) wake-storm on every @spawn (already labeled backport 1.13; tasks: only switch to the scheduler task when the current task is done, so interrupts reach user code #62298/scheduler: fall back to a pool wake when the targeted thread wake is a no-op #62371 build on it)
• partr: wake task's owner thread when skipping it in scheduler #62318 — partr: wake task's owner thread when skipping it in scheduler
• tasks: only switch to the scheduler task when the current task is done, so interrupts reach user code #62298 — tasks: only switch to the scheduler task when the current task is done, so interrupts reach user code
• scheduler: fall back to a pool wake when the targeted thread wake is a no-op #62371 — scheduler: fall back to a pool wake when the targeted thread wake is a no-op
• test: synchronize interrupt tests on observable state, not fixed sleeps #62385 — test: synchronize interrupt tests on observable state, not fixed sleeps
• test/misc: remove flaky Distributed interrupt test #62422 — test/misc: remove flaky Distributed interrupt test
• interrupting: Make asynchronous interrupts survivable at the REPL #62471 — interrupting: Make asynchronous interrupts survivable at the REPL
• scheduler: handle enqueuing tasks into threadpools with zero threads #62389 — scheduler: handle enqueuing tasks into threadpools with zero threads (not interrupt-specific, but fixes a hard crash in the enq_work path modified by tasks: only switch to the scheduler task when the current task is done, so interrupts reach user code #62298/scheduler: fall back to a pool wake when the targeted thread wake is a no-op #62371)
• [Windows] Use TimerQueue API to register Profile watchdog #60463 — [Windows] Use TimerQueue API to register Profile watchdog
• [Windows] Add DLL notification queue to avoid deadlock #60690 — [Windows] Add DLL notification queue to avoid deadlock

The last two complete the Windows profiler-deadlock fix chain (#60056 is already on this branch): with the scheduler task relanded, the Profile debuginfo-registration test deadlocks on Windows CI exactly as it did on master before these fixes (#60042, #60306).

#62372 was also considered but is already on this branch via an earlier backport.

Backport adaptations

• Fix interrupting. Add tests. #62069, gf.c: master needed a sigatomic guard around jl_type_infer because Add function to emit multiple CodeInstances to the JIT atomically (remove jl_typeinf_lock) #61255 removed jl_typeinf_lock; 1.13 still takes that lock (the Temporarily reintroduce a global type inference lock #60689 backport), which already defers signals for its duration, so the guard is dropped in a separate commit on top of the faithful cherry-pick.
• partr: wake task's owner thread when skipping it in scheduler #62318/scheduler: fall back to a pool wake when the targeted thread wake is a no-op #62371, partr.jl: master's multiq_deletemin was restructured by the labeled-break syntax PR syntax: Add labeled block break #60481, which can't come to 1.13; the hunks are placed into 1.13's @goto retry structure instead.
• interrupting: Make asynchronous interrupts survivable at the REPL #62471, signals-mach.c: master's mach_safepoint_trampoline (from [macOS] Handle GC safepoint on-thread #61341, not backported) doesn't exist on 1.13; the eh == NULL deferral guard is applied to 1.13's equivalent decision point in catch_mach_exception_raise.
• interrupting: Make asynchronous interrupts survivable at the REPL #62471, task dispatcher: adapted to src/llvm-julia-task-dispatcher.h (pre-Add function to emit multiple CodeInstances to the JIT atomically (remove jl_typeinf_lock) #61255 name/structure): same dispatcher_sigdefer_guard, placed around the inline-run/drain path and work_until.
• interrupting: Make asynchronous interrupts survivable at the REPL #62471, REPL.jl: parse_repl_input_line doesn't exist on 1.13; the BasicREPL loop keeps Base.parse_input_line.
• [Windows] Add DLL notification queue to avoid deadlock #60690, stackwalk.c: 1.13's jl_unw_init keeps its (void)from_signal_handler; line (master had dropped the parameter); the Fix windows profiler deadlock #60056 test-enablement hunk reduces to nothing since master later re-skipped that stress test.
• NEWS.md hunks were dropped.