[feature] Emit diagnostics when codecs drop request fields, and on the stream path #1215

Problem

Several request fields are removed during translation without a TranslationDiagnostic, so
callers cannot tell that the upstream received a weaker request. #607 made the server surface
diagnostics that codecs already emit; this issue is about codecs that emit none for these
drops, and about the stream path, which has no diagnostics channel.

Reproduction on main at f60a7a22, TranslationEngine::default() and
TranslationPolicy::default(), translate_request("anthropic_messages", target, …):

{"model":"m","max_tokens":100,"top_k":5,
 "system":[{"type":"text","text":"sys","cache_control":{"type":"ephemeral"}}],
 "messages":[{"role":"user","content":[
   {"type":"text","text":"hi","cache_control":{"type":"ephemeral"}},
   {"type":"document","source":{"type":"text","media_type":"text/plain","data":"doc"},
    "citations":{"enabled":true}}]}]}

TranslationEngine::translate_event returns Vec<Value> only, so losses found mid-stream
(for example citations, redacted_thinking or server-tool blocks with no target
representation) cannot be reported either.

Proposal

1. Emit a warning diagnostic with a JSON path for each known-unrepresentable request field per
   codec pair, starting with the four above. lossy_conversion_policy: Reject would then reject
   them, as it already does for provider tools.
2. Add a diagnostics sink to StreamTranslationState (for example a
   Vec<TranslationDiagnostic> the caller drains), filled by stream codecs when they drop or
   degrade content.

Happy to send a PR for (1) if this shape looks right.