Tag evals with the tracer version and ml app - #12184

What does this PR do?

Adds the ddtrace.version and ml_app tags to every submitted evaluation metric, so an eval can be filtered by tag and not only through the top level ml_app payload field.

Motivation

Spotted while reviewing #12129: feedback events carry those two tags but evals never did, making the Java tracer the odd one out.

• dd-trace-js builds evaluation and feedback tags with the same buildMetricTags(tags, mlApp, kind) (packages/dd-trace/src/llmobs/eval-metric.js), which always prefixes ddtrace.version and ml_app.
• dd-trace-py starts both evaluation_tags and feedback_tags from {"ddtrace.version": __version__, "ml_app": ml_app} (ddtrace/llmobs/_eval_metric.py).

Stacked on top of #12129, whose shared IntakeTags.flatten helper this reuses. Review that one first.

Wire change

Purely additive, but note the tags field is now always present in the eval payload — it used to be omitted entirely when the caller passed no tags.

 {
   "event_kind": "evaluation",
   "ml_app": "my-app",
-  "tags": ["source:web-ui"]
+  "tags": ["ddtrace.version:1.2.3", "ml_app:my-app", "source:web-ui"]
 }

Testing

LLMObsEvalTest, which pins the payload key set, is updated to cover both cases: user supplied tags, and none.