Add Zipkin Virtual GenAI e2e test with zipkin_json exporter

[wu-sheng]

Add Zipkin Virtual GenAI e2e test

This is an alternative to #13781 (cc @peachisai). It adds the same Zipkin Virtual GenAI e2e test but fixes the pip dependency conflict properly.

Root cause: why single vs multiple pip install matters

The original PR #13781 used separate pip install commands (one per package) to work around a build failure. Here's why:

Single pip install (all packages in one command) — pip resolves all dependency constraints simultaneously and fails with ResolutionImpossible because:

• opentelemetry-exporter-zipkin-proto-http requires protobuf ~= 3.12 (i.e., >=3.12, <4.0)
• opentelemetry-proto (pulled by opentelemetry-exporter-otlp-proto-grpc) requires protobuf >= 5.0

These two constraints are mathematically irreconcilable — no protobuf version can satisfy both.

Multiple pip install (separate commands) — pip resolves each package independently, and later installs silently overwrite transitive dependencies from earlier installs. This "succeeds" but creates a broken environment where pip check reports violations:

• opentelemetry-proto requires protobuf>=5.0, but protobuf==3.20.3 is installed
• googleapis-common-protos requires protobuf>=4.25.8, but protobuf==3.20.3 is installed

The split install only works at runtime by accident — the zipkin code path doesn't import from opentelemetry-proto.

Fix

Use OTEL_TRACES_EXPORTER=zipkin_json instead of zipkin. The zipkin entry point maps to opentelemetry-exporter-zipkin-proto-http (protobuf format), while zipkin_json maps to opentelemetry-exporter-zipkin-json (JSON format) which has no protobuf dependency.

• Remove opentelemetry-exporter-zipkin (meta-package) and opentelemetry-exporter-zipkin-proto-http from Dockerfile
• Keep opentelemetry-exporter-zipkin-json — single pip install, clean dependencies, pip check passes
• SkyWalking's zipkin receiver accepts both JSON and protobuf on the same /api/v2/spans endpoint

Locally verified end-to-end: OAP with BanyanDB receives spans via zipkin_json, creates VIRTUAL_GENAI service (openai), registers model instance (gpt-4.1-mini-2025-04-14), and populates all GenAI metrics (CPM, SLA, response time, token counts, estimated cost).

• Tests(including UT, IT, E2E) are added to verify the new feature.
• Update the CHANGES log.

[wu-sheng]

@peachisai This should work, let's see.