fix(llm, cli, anthropic): Bound provider rebuild loop and heal turns

[JeanMertz]

Anthropic rejects a thinking block whenever any thinking or redacted-thinking block in the latest assistant message was rewritten, not only when a signature goes stale. The repair now recognizes both rejection messages and, when the offending block sits in the latest assistant turn, downgrades every thinking block in that message in one patch set instead of one block per round. Anthropic requires that message's thinking blocks to come back unchanged as a whole, so patching one block at a time left the message half-rewritten and the next request failed the same way. A conversation already left in that half-rewritten state is now healed by the same patch. Earlier turns keep their narrower, single-block repair since they carry no such constraint.

A provider that keeps asking to rebuild the request without actually fixing anything could previously loop forever, resending the full conversation each time. jp query now refuses a rebuild unless the preceding patch changed the stream in a way guaranteed to make progress (PatchAction::shrinks_stream), and caps consecutive rebuilds within a turn at 10. Hitting either limit aborts the turn with an error instead of looping, after flushing whatever was already streamed so no partial response is lost. The counter resets once a turn streams a successful event, so later, unrelated repairs are not penalized by an earlier one.