feat(otel): GenAI semconv metrics and richer spans (ellmer port)

CHANGELOG.md

@@ -13,6 +13,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 ### New features
 
 * Providers now expose a `base_url` attribute (`chat.provider.base_url`) reporting the endpoint requests are sent to, when the provider can determine it. This is groundwork for richer telemetry (e.g. OpenTelemetry `server.address`/`server.port` attributes) and is handy for confirming which endpoint a provider resolved.
+* The OpenTelemetry integration now emits **metrics** alongside spans, following the [GenAI semantic conventions](https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-metrics/): per-request inference duration and token-usage histograms (`gen_ai.client.inference.*`), cumulative token-usage counters (including cache-read and reasoning tokens, broken down by `gen_ai.token.modality`), streaming latency (`gen_ai.client.inference.time_to_first_chunk`, `gen_ai.client.inference.time_per_output_chunk`), tool execution duration (`gen_ai.execute_tool.duration`), and per-invocation agent metrics (`gen_ai.invoke_agent.duration`, `gen_ai.invoke_agent.inference_calls`, `gen_ai.invoke_agent.tool_calls`). Metrics flow through the same SDK configuration as traces, so no extra setup is needed.
+* OpenTelemetry spans are richer and more accurate. Chat spans now record `server.address`/`server.port`, the request parameters actually sent (`gen_ai.request.*`, including `gen_ai.request.stream` and reasoning effort), the finish reason (`gen_ai.response.finish_reasons`), cache-read and reasoning token counts, time to first chunk, and (with content capture enabled) the tool definitions sent with the request. The `invoke_agent` span is now an `internal`-kind span and records duration, inference/tool call counts, and total token usage. Successful requests are now marked with an `OK` span status (previously unset), and failed, cancelled (`StreamController.cancel()`), and interrupted (e.g. abandoned streams) requests now report an error status and `error.type` on both the spans and the duration metrics, matching ellmer.
 * The result of `chat.files.get()` (and elements of `chat.files.list()`) can now be passed straight to `.chat()` (and anywhere else content is accepted), so a file uploaded in an earlier session can be reused from its id alone: `chat.chat("Summarize this.", chat.files.get("file_011CRabc..."))`. When the provider doesn't report a MIME type (OpenAI's Files API never does), it's guessed from the filename. For when you want the underlying `ContentUploaded` reference itself (e.g. to build a `UserTurn` by hand), `FileMetadata` gains a `.to_content()` method. This works via a new general convention: any object implementing `to_content() -> Content` (see `chatlas.types.SupportsContent`) is accepted anywhere content is passed, so third-party packages can make their own objects usable as chat inputs. (#417)
 
 ### Changes