Log the feedback a model drafts about its own turn

[mipsel64]

Claude Code's SendFeedback tool drafts a bug report when it judges that a turn
went wrong, and a model behind tinyllm fires it at its own mistakes. Those drafts
are also a bug report about tinyllm — a model saying it misread a request may be
describing something the translation dropped: a message part, a tool result,
reasoning lost across a continuation. They were queued in the client and never
reached the operator running the gateway.

Changes

• RequestBody::drafted_feedback reads the tool call out of the request that
  carries it; endpoint::run logs the full input at warn with the model and
  request ID.
• It inspects only the last exchange: an assistant message holding a
  SendFeedback tool_use, followed by the user message with its results. That
  matches exactly one request per draft, so a stateless gateway reports each once
  instead of re-logging it every turn as the history grows.
• Documented in README.md. No new configuration.

Why not block the tool

Stripping the definition from the forwarded tool list would stop the drafts, but
it would also stop the signal, and the call is a noop for the session: the draft
is queued in the client, which is why it never interrupts the conversation.

Swallowing the tool_use in the outbound response would silence it without
losing the log, but dropping a call that was the model's only action ends the
turn early and stops work mid-task — and a stream cannot tell that in advance
without buffering.

Testing

drafted_feedback_is_read_from_the_turn_that_just_completed covers the carrying
request, including a draft batched with another tool call, and the six shapes
that must stay silent — among them the same draft further back in the history.
Full suite: 112 passed, clippy clean.