fix(cost): align GLM 5.2 pricing with the benchmark catalog; record judge calls in session stats (#325)

* fix(cost): align GLM 5.2 reference pricing with the benchmark catalog

The glm-5.2-fp8 entry carried SiliconFlow's serverless list price, while
the benchmark pricing catalog prices zai-org/glm-5.2 from a different
reference tier, so cost reports computed from this table disagreed with
the benchmark dashboards for the same traffic. Adopt the catalog rates
and add the hub-facing zai-org/glm-5.2 id variants, which previously had
no entry at all and therefore priced hub GLM traffic at zero via the
unknown-model fallback.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Signed-off-by: Lin Jia <linj@nvidia.com>

* fix(cost): keep the GLM pricing entry to the local deployment id

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Signed-off-by: Lin Jia <linj@nvidia.com>

* feat(server): record judge calls in per-session routing stats

Judge and classifier calls were recorded only in the server-wide
classifier stats bucket, so per-session routing snapshots reported the
routed tiers but silently omitted judge token overhead — sessions
looked cheaper than they were, and the overhead was unrecoverable once
the server shut down. The stats observer now also appends successful
non-routed calls to the routing log under the judge's model id with a
'classifier' tier tag. Routed calls stay off this path because the
served response is already logged with terminal usage in
usage_metrics::observe.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Signed-off-by: Lin Jia <linj@nvidia.com>

* docs(cost): clarify GLM pricing provenance

Signed-off-by: Lin Jia <linj@nvidia.com>

---------

Signed-off-by: Lin Jia <linj@nvidia.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>