feat: raise default sizes for non-evictable shared dicts

[nic-6443]

Description

The prometheus-metrics, service-discovery (kubernetes / consul / nacos), and tracing_buffer shared dicts reject writes and drop data when full — unlike caches such as plugin-limit-count or balancer-ewma, they have no self-healing eviction. Their current defaults overflow on medium and large deployments, producing symptoms that are hard to diagnose: silently missing Prometheus series, upstream nodes that never resolve, or lost trace spans.

This raises the defaults so a typical deployment works without hand-tuning lua_shared_dict / shared_size:

Dict	Old	New	Stores	Scales with
prometheus-metrics	15m (meta) / 10m (http)	128m	metric series	route × label × bucket cardinality
kubernetes / consul shared_size	1m	64m	discovered endpoints	services × nodes
nacos / nacos-stream	10m	64m	discovered endpoints	services × nodes
tracing_buffer	10m	32m	span backlog awaiting report	in-flight span volume

Backward compatible. An explicit size in config.yaml (or a discovery config's shared_size) always overrides the default, so deployments that already tune these are unaffected.

Trade-off. These are larger reserved zones. On Linux the memory is committed lazily as a dict actually fills, so a small/idle deployment does not pay the full cost up front, and deployments that don't use a subsystem (discovery / tracing) can lower or omit the value. Feedback on the exact defaults is welcome — the intent is to pick values that cover the majority of deployments rather than the smallest possible footprint.

Checklist

• Docs updated (kubernetes discovery shared_size default, en + zh)
• Tests updated (CLI render + kubernetes discovery default assertions)
• Verified locally that the rendered nginx.conf reflects the new defaults