feat(adk): surface prompt-cache hit metrics via genai cached content token count

[MartinForReal]

Summary

Closes #2669 — surfaces provider-reported prompt-cache hit / cached-token counts from the kagent model adapters into genai.GenerateContentResponseUsageMetadata.CachedContentTokenCount, so cached prompt-token spend and prompt-cache hit rate are observable via A2A task usage and the llm_response trace attributes instead of being silently dropped.

Adapter changes

Populate CachedContentTokenCount (int32) at each adapter, mapping the provider's cached-prompt-token counter; defaults to 0 when the provider omits the field (nil-safe):

• OpenAI Responses — go/adk/pkg/models/openai_responses.go: ResponseUsage.InputTokensDetails.CachedTokens
• OpenAI Chat (streaming + non-streaming) — go/adk/pkg/models/openai_adk.go: PromptTokensDetails.CachedTokens
• Anthropic (streaming + non-streaming) — go/adk/pkg/models/anthropic_adk.go: Usage.CacheReadInputTokens
• Amazon Bedrock Converse (streaming + non-streaming) — go/adk/pkg/models/bedrock.go: Usage.CacheReadInputTokens
• Python OpenAI runtime — python/packages/kagent-adk/src/kagent/adk/models/_openai.py: prompt_tokens_details.cached_tokens

This is gap 1 ("populate the cached-token counts in the adapters") from #2669. Emitting a dedicated gen_ai.token.type=cached metric remains the separately-tracked follow-up surfaced in #2148/#2149, and upstream ADK already records gen_ai.usage.cache_read.input_tokens once CachedContentTokenCount is non-zero.

Tests

• Go unit tests added/updated for the cached-token mapping:
  • anthropic_adk_test.go — new TestAnthropicNonStreamingCachedTokens
  • openai_responses_test.go — responses non-streaming, GenerateContent, and streaming cached-token assertions
  • openai_adk_test.go — chat completion cached-token assertion
• Python tests/unittests/models/test_openai.py — asserts cached_content_token_count is populated and defaults to 0 when absent.

Validation

• go build ./... ✅
• go vet ./... ✅
• go test ./adk/pkg/... ✅ (note: go/adk/pkg/session shows a pre-existing Windows-only TempDir lockfile cleanup failure unrelated to this change; all models package tests pass)
• gofmt -l on changed Go files: empty ✅
• Python unit tests for the changed file could not be executed in this workspace because dependency installation failed on repeated files.pythonhosted.org TLS (HandshakeFailure) downloads — an environment/network issue, not a code defect (recorded in notes; CI will exercise them).

This branch is DCO-signed (git commit -s). Not a draft.