When a replica node is taken offline, it will briefly respond to
commands with some form of `CLUSTERDOWN` error. Presently while this did
cause PhpRedis to invalidate the slot cache, we did not immediately
remap the topology.

What this meant, is that for an already connected cluster, we might keep
sending readonly commands to this now decomissioned cluster replica for
the duration of the request.

Additionally, if the request was long lived enough we may still have it
mapped if the replica was brought back online at which point the client
might see dozens of `LOADING` exceptions.

This commit just changes our behavior when we detect any `CLUSTERDOWN`
error such that we flush the slot cache, disconnect the particular node,
and also attempt a remap of the topology.

This commit adds a new `ini` setting `redis.clusters.slot_cache_expiry`
which specifies in seconds how long a cached slot map should live before
being refreshed. The setting is disabled by default to avoid breaking
backward compatibility.

This commit adds a mechanism to invalidate every `RedisCluster` slot
cache across all forked workers with one call to any of the workers.

The invalidation itslef works by using a `uint64_t` generation counter
which is allocated in shared memory on `MINIT` which is stored in every
slot cache when it is created.

When a user wants to invalidate any of the slot caches they can call
`RedisCluster::invalidateSlotCaches()` which simply does an atomic
increment on this global genertion value.

When a slot cache is retreived from the persistent list, the internal
generation is compared to the global shared memory generation and if
they don't match, the cache is not used, which is the same as
invalidting it as RedisCluster will write a new slot cache on object
destruction.

The feature is only compiled if `config.m4` determines that using `mmap`
to allocate shared memory is available and works.

Additionally, the feature must be explicitly enabled by setting the
`redis.clusters.shared_slot_cache_invalidation`.

`time(NULL)` isn't monotonic and can run backwards so switch to
`zend_hrtime` which uses `clock_gettime(CLOCK_MONOTONIC)` on Linux and
`QueryPerformanceCounter`on Windows.