Fix Eval trace structure inconsistencies

[David Elner (delner)]

There are numerous issues with the Eval trace structure when compared to the Java SDK (the other OTel-based implementation) Python, and TypeScript traces.

Changes

Issue	Before	After
Single shared score span	All scorers ran inside one "score" span with all scores aggregated on it	Each scorer gets its own "score" span as a direct child of the eval span, matching Java/Python/TS
Missing purpose: "scorer" on score spans	span_attributes only had {type: "score"}	span_attributes includes {type: "score", name: scorer_name, purpose: "scorer"} — used by the platform to filter scorer spans from cost/latency calculations
Missing scorer input/output on score spans	Score span had no input_json or output_json	Each score span logs input_json (input, expected, output, metadata) and output_json (scores hash), matching Python/TS expected output
Eval span input_json not wrapped	Raw value (e.g., "hello")	Wrapped as {input: "hello"}, matching Java SDK
Eval span output_json not wrapped	Raw value (e.g., "HELLO")	Wrapped as {output: "HELLO"}, matching Java SDK
Missing metadata on eval span	Case metadata not logged on the eval span	Case metadata set as braintrust.metadata on the eval span, matching Java SDK
Missing output_json on eval span when task errors	No output_json attribute set at all	Sets {output: null}, matching Java SDK
Eval span attributes missing on task error	span_attributes, input_json, expected, metadata, origin were set after task+scorers, so they were skipped on task error	All known attributes set before task execution so they're present regardless of task outcome
Eval spans not isolated from ambient trace context	Used tracer.in_span("eval") which inherits any active parent span (e.g., a Sidekiq job span)	Uses tracer.start_root_span("eval") so each eval case starts its own independent trace, matching Java's setNoParent()