feat(adk): emit cached token series on gen_ai token usage metric (#2669)

[MartinForReal]

Step 2 of #2669: emit a cached token-type series on the existing gen_ai.client.token.usage metric so prompt-cache hits become aggregatable (gen_ai_token_type="cached").

This change stacks on top of the token-usage histogram introduced in (#2673) — "feat(adk): expose gen_ai token usage Prometheus metric" — which must merge first.

What changes

• Adds a cached series (gen_ai_token_type="cached") to the gen_ai_client_token_usage histogram.
• RecordTokenUsage records the cached prompt-token count when positive; zero/negative counts are skipped, matching the existing input/output behavior.
• recordTokenUsage in the A2A executor wires UsageMetadata.CachedContentTokenCount through as the cached token count.

Testing

• Unit tests cover the cached series (recorded when enabled and > 0, skipped when disabled or non-positive) and the executor wiring (a non-partial event with CachedContentTokenCount emits a gen_ai_token_type="cached" series valued at the cached count; partial events and absent/zero cached counts emit none).
• gofmt -l clean on all changed files; go build ./..., go vet ./adk/..., go test ./adk/pkg/a2a/... and go test ./adk/pkg/telemetry/... pass.

DCO

Signed-off-by: MartinForReal fanshangxiang@gmail.com

Refs: #2669, #2673