fix(cli, llm, serve-web): Join split reasoning into one region

[JeanMertz]

Anthropic interrupts a thinking block with an opaque redacted_thinking block and resumes the thinking in a third block, splitting one region of reasoning mid-word across three items. Each thinking block carries its own signature, and a ConversationEvent holds one signature field, so the three items cannot be merged upstream without dropping a signature and failing the next request's signature validation.

Rendering previously treated every ChatResponse boundary as a block boundary, which broke this text apart on screen, in jp c print replay, and in the web viewer. Text-bearing responses (Message, Reasoning) now concatenate across consecutive events of the same kind, and the visual break between segments comes from blank lines in the stored content instead of from the response boundary. Discrete responses (Structured) are unaffected: each one still closes its json fence at the response boundary, since a JSON value cannot encode its own framing.

Reasoning events are now persisted even when their text is empty, because the attached metadata (a redacted-thinking payload or a thinking signature) can be the only reason the event exists; dropping whitespace-only reasoning would lose that metadata. The web viewer's render_events gained a small buffer that accumulates consecutive text events of one kind before parsing them as markdown, so it stays consistent with the CLI's rendering.

jp_llm's MultiProviderMetadata now derives its Map conversion from Serialize instead of building the map by hand, since the struct fields already match the metadata keys one for one.