2.0.0 — Gemma text-call parsing moves to AIChatKitMLX; failed turns keep role alternation

Gemma text tool-call parsing moves out of AIChatUI into AIChatKitMLX, and a turn that fails before
any output no longer leaves two user messages in a row.

Breaking

• Removed GemmaOutputRecovery, GemmaToolArguments and EmbeddedToolCallParser (all were public
  in AIChatUI).
• ChatSession no longer parses tool calls out of assistant text (call:name{…},
  <tool_call>{json}</tool_call>, ```tool_code). Text a provider streams as text is shown as text.
  A provider that talks to a model which writes calls as text must turn them into .toolCallComplete
  events itself. AIChatKitMLX 1.4.0+ does this for Gemma, at stream level.
• ChatSession no longer rewrites tool-call arguments. It trims them, treats blank as {}, and still
  sends anything that is not a JSON object to the provider as {}.

Fixed

• A turn that ends in an error, or an empty reply, before any output is now dropped from the history
  sent to the provider, the same way a cancel before output is. Previously the next send could send
  two consecutive user turns, which strict chat templates (Gemma) reject. Errors after partial
  output are unchanged: the partial reply is committed, so roles already alternate.

Added

• ChatSession.UserEntry.isFailed (default false): set on such a turn. ConversationView shows it
  dimmed with the caption “Not answered — the model won't see this message” and a matching VoiceOver
  label (UserMessagePresentation.failedCaption).
• docs/CHAT_SESSION_BEHAVIOUR.md documents the thinking tile and where tool-call parsing lives.

Changed

• README: “Known limits” is now “Upgrade notes”. Swift / platform / license badges.

Upgrading from 1.x

1. Using AIChatKitMLX with Gemma? Update it to 1.4.0 or later in the same change. With an older
   AIChatKitMLX, a Gemma model that writes <tool_call> XML or tool_code blocks would show them as
   plain text, because 2.0.0 no longer recovers them in the UI.
2. Called GemmaOutputRecovery, GemmaToolArguments or EmbeddedToolCallParser directly? They are
   gone. Use a provider that emits .toolCallComplete (AIChatKitMLX 1.4.0+ for Gemma), or copy the
   parser into your own provider.
3. Draw your own user-message rows? Treat UserEntry.isFailed like isCancelled: the message was
   never answered and is not in the provider history. Resending it retries.
4. Depend on other NerdSnipe packages? They must allow AIChatKit 2.x. Compatible releases:
   AIChatKitMLX 1.4.0, AiPersona 1.1.1, GHLAIService 0.2.1. AIChatKitLlama 1.0.2 and earlier require
   AIChatKit <2.0.0.
5. Nothing else in the public API changed.