[Feature] upstream: defer DNS resolution for unreachable hosts

[Feature] upstream: defer DNS resolution for unreachable hosts

Previously rspamd_upstreams_add_upstream() returned FALSE whenever
rspamd_parse_host_port_priority() could not resolve a hostname, so a
single DNS hiccup at config time would drop the upstream list and
cascade into module init failures (issue #6000 was one symptom).

Introduce RSPAMD_UPSTREAM_FLAG_PENDING_RESOLVE: when DNS fails for a
hostname-style input we now keep the parsed host and port on the
upstream, mark it pending, and let the existing async lazy-resolve
machinery retry. Pending upstreams are deliberately kept out of the
`alive` list so existing selectors (round-robin, hashed, master/slave)
and the ring/heap rotators do not need to learn a new state - they
continue to see only usable upstreams. The probe-mode fallback that
walks `ups` directly skips pending entries explicitly.

set_active() schedules a fast initial resolve (1s with jitter) for
pending upstreams; lazy_resolve_cb() backs off exponentially up to
60s while the upstream stays pending. update_addrs() handles the
empty-initial case by reading the port from the stashed
`deferred_port` field, and on success rspamd_upstream_promote_pending
clears the flag and inserts the upstream into `alive` (initialising
token-bucket state if needed) and fires the WATCH_ONLINE event.

This changes the failure mode for every consumer of upstreams (Redis,
ClamAV, ICAP, ClickHouse, ...): a misconfigured or briefly-down DNS
no longer prevents the daemon from starting, and recovers automa-
tically without a restart. Callers that pick from an alive-empty list
already had to handle nil from rspamd_upstream_get; later commits
audit and tighten the Lua callers that did not.