timer/migration: Fix quick check reporting late expiry

When a CPU is the last active in the hierarchy and it tries to enter
into idle, the quick check looking up the next event towards cpuidle
heuristics may report a too late expiry, such as in the following
scenario:

                        [GRP1:0]
                     migrator = NONE
                     active   = NONE
                     nextevt  = T0:0, T0:1
                     /              \
          [GRP0:0]                  [GRP0:1]
       migrator = NONE           migrator = NONE
       active   = NONE           active   = NONE
       nextevt  = T0, T1         nextevt  = T2
       /         \                /         \
      0           1              2           3
    idle       idle           idle         idle

0) The whole system is idle, and CPU 0 was the last migrator. CPU 0 has
a timer (T0), CPU 1 has a timer (T1) and CPU 2 has a timer (T2). The
expire order is T0 < T1 < T2.

                        [GRP1:0]
                     migrator = GRP0:0
                     active   = GRP0:0
                     nextevt  = T0:0(i), T0:1
                   /              \
          [GRP0:0]                  [GRP0:1]
       migrator = CPU0           migrator = NONE
       active   = CPU0           active   = NONE
       nextevt  = T0(i), T1      nextevt  = T2
       /         \                /         \
      0           1              2           3
    active       idle           idle         idle

1) CPU 0 becomes active. The (i) means a now ignored timer.

                        [GRP1:0]
                     migrator = GRP0:0
                     active   = GRP0:0
                     nextevt  = T0:1
                     /              \
          [GRP0:0]                  [GRP0:1]
       migrator = CPU0           migrator = NONE
       active   = CPU0           active   = NONE
       nextevt  = T1             nextevt  = T2
       /         \                /         \
      0           1              2           3
    active       idle           idle         idle

2) CPU 0 handles remote. No timer actually expired but ignored timers
   have been cleaned out and their sibling's timers haven't been
   propagated. As a result the top level's next event is T2 and not T1.

3) CPU 0 tries to enter idle without any global timer enqueued and calls
   tmigr_quick_check(). The expiry of T2 is returned instead of the
   expiry of T1.

When the quick check returns an expiry that is too late, the cpuidle
governor may pick up a C-state that is too deep. This may be result into
undesired CPU wake up latency if the next timer is actually close enough.

Fix this with assuming that expiries aren't sorted top-down while
performing the quick check. Pick up instead the earliest encountered one
while walking up the hierarchy.

7ee98877 ("timers: Implement the hierarchical pull model")
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://lore.kernel.org/r/20240305002822.18130-1-frederic@kernel.org