feat(telemetry): emit the routing decision as span attributes

[Bekhouche]

Summary

Mirror a collection.router decision onto the request's inference span. Spans
now carry llm.route.{collection,to,matched_rule,default_used} under the
existing llm.* namespace, plus the decision's estimated_cost (from #2763)
flattened into llm.route.cost_*. Previously the decision was visible only to
the client on x_lemonade_route, so a trace showed which model ran but not that
a policy picked it, or why.

The span is created inside Router, where the Decision is out of scope, so
the decision travels on a thread-local published by Server — the same pattern
as the existing trace-context fields, including the per-request clear that stops
a stale value leaking onto a later request on a reused worker. It is published
from route_collection_request, the one function every dispatch path already
goes through, so chat, completions and responses cannot drift apart.

Distinct from the routing telemetry in #2968, which counts decisions and
switches server-wide; this is per-request detail on the span itself.

Note the cost half stays empty for local models today: nothing populates cost
metadata for them, since USER_DEFINED_MODEL_PROPS drops cost_tier /
cost_input_per_million / cost_output_per_million on registration. The
routing fields are unaffected.

Scope

telemetry.h / telemetry.cpp — the thread-local, plus route_span_attributes
(payload -> attributes, a pure function) and apply_route_attributes.
server.cpp — per-request clear, and publishing the decision at the dispatch
choke point. router.cpp — one call in each of the five span sites (chat,
completions and responses, streaming and not). test_telemetry_helpers.cpp.

Testing

C++ build + full cpp-ci CTest suite, 69/69. Twelve new cases cover the
mapping: an empty, malformed or non-object payload emits nothing, a nested
member is skipped rather than stringified, and author-set estimated_cost keys
cannot shadow the decision's own.

Also confirmed on a live server rather than only in unit tests. Registered a
router collection over two local GGUF candidates, connected to the spans
WebSocket, and captured the emitted span for a routed request:

SPAN name=chat.completions
llm.backend            = llamacpp
gen_ai.request.model   = Qwen3-1.7B-GGUF
llm.route.collection   = user.Router-Demo
llm.route.matched_rule = code-to-bigger
llm.route.to           = Qwen3-1.7B-GGUF
llm.route.default_used = False

Documentation

No docs change.

Breaking Changes

None. Purely additive attributes on an existing span; a request that is not
routed emits nothing new.

AI-assisted contribution

Yes.