Pass the permit turn on when a woken waiter cannot start (#197)

[gnacho]

Closes #197.

Problem

Since the ETag gate landed, the tray icon stays on cloud-sync forever even when every folder is idle: the aggregate reports a SyncQueued/Syncing folder that never runs (no nextcloudcmd process, ~0 CPU).

Root cause

SyncPermit::release() wakes exactly ONE waiter and assumes it will try_acquire again. But the woken scheduler's start() can bail out through an early guard (empty queue, already running, paused, offline, auth/delete gates...) without taking the permit and without waking the next waiter. The turn is swallowed: the permit sits free while every other waiter sleeps forever.

With the pre-gate full reconciliations, runs took minutes and the interleavings were rare. The ETag-skip makes runs milliseconds long, so the double-waiter scenario happens routinely: a folder that receives a second trigger while queued registers two waiters; the leftover waiter's abort strands the folders behind it in SyncQueued forever.

Fix

start() now reports a StartAttempt (Ran / Waiting / Aborted). The permit-waiter callback checks the attempt: when it aborts (nothing launched, nothing waiting), it calls permit.release() to pass the turn to the next waiter (release tolerates in_use == 0). Waiting (the scheduler re-queued itself) and Ran keep the turn, exactly as before.

Testing

New regression test aborted_permit_waiter_passes_the_turn_to_the_next_folder (three schedulers on one permit, double waiter on the middle one), red-green verified: without the propagation third never starts and the assertion fails; with it every folder ends IdleOk.

Gate: 706 tests, clippy clean, fmt applied.