Dynamo router total per-request routing overhead histogram

The routing_overhead family is a per-phase decomposition of the ROUTER'S OWN CPU WORK: dynamo_router_overhead_block_hashing_ms, overhead_indexer_find_matches_ms, overhead_seq_hashing_ms, overhead_scheduling_ms and this roll-up. Two important corrections a reader would otherwise get wrong. (1) THE NAME PREFIX is dynamo_router (name_prefix::ROUTER), not dynamo_routing_overhead - prometheus_names.py has a ROUTING_OVERHEAD constant holding the SUFFIXES, which is an easy trap since it looks like a prefix. (2) THERE ARE TWO LADDERS in lib/llm/src/kv_router/metrics.rs: compute_overhead_buckets() = prometheus::exponential_buckets(0.001, 2.0, 15) is used for the two pure-CPU phases (block hashing, sequence hashing), while async_overhead_buckets() = exponential_buckets(0.01, 3.0, 17) is used for indexer find_matches, scheduling, TOTAL and shared-cache query - this record carries the async ladder because that is the one on the total. Note the units are MILLISECONDS while every surrounding latency metric is in seconds: a query that forgets the suffix is wrong by 1000x and silently so. Every histogram in the family carries a const_label router_id (the router's etcd discovery.instance_id), so series are per-router not fleet-wide. The doc comment on register() notes this is NOT available on the standalone router, which has no frontend HTTP server - a real absence that looks like a config error.