fix(dictation): stop leaking raw <think> content when stripped output is empty

[postoso]

What

In the non-streaming LLM response parsers, content fell back to the RAW content when the thinking-stripped content was empty (cleanedContent.isEmpty ? rawContent : cleanedContent). When a reply is all thinking (<think>...</think> with nothing after), cleanedContent is empty, so the raw tagged text leaked downstream and got inserted into the user's document. The streaming path already returns empty (which trips the existing emptyResponse error), so the two paths disagreed.

How

Return cleanedContent directly at both parser sites (parseResponsesResponse, parseMessageResponse). An all-thinking reply now yields empty content and trips the existing empty-response handling instead of leaking the tags, matching the streaming path. The two parsers are widened from private to internal so the new tests can reach them; behavior is otherwise unchanged.

Tests

New LLMClientThinkingTagTests (5 methods): all-thinking yields empty, content after </think> is kept clean, plain content is unchanged, covering both the chat-completions and Responses API parsers. xcodebuild test is green and swiftlint --strict is clean.