feat(openai): keep a builtin tool call's status and mark failed calls as error spans (#991)

The Responses API reports each builtin output item (`web_search_call`,
`file_search_call`, `code_interpreter_call`) with a `status` — `completed`,
`failed` or `incomplete`. Both decode paths dropped it together with the
item's id/type, so a failed web search and one that found nothing were
indistinguishable downstream: the OTel `execute_tool` child span was
always `:ok` with no result.

The status now rides on the tool call's metadata (kept off the arguments,
which the item does not replay anyway) on the streaming and the buffered
path, and `ReqLLM.Telemetry.OpenTelemetry.tool_spans/2` turns `failed` /
`incomplete` into an error span stub carrying `error.type`, which the
translator applies to the child span as usual.


Claude-Session: https://claude.ai/code/session_01XUXfaDAAeFok1GCCULzyVD

Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>