perf: skip syntax highlighting while a message is streaming

[k0n9-devx]

Problem

While a message streams, every message_update re-renders the live MessageView, and CodeBlock re-runs Prism over the entire growing code fence on each update. A long code block gets re-tokenized hundreds of times before it completes. On phones this is the largest single CPU cost of streamed rendering — enough to cause visible thermal throttling on remote access.

Two smaller issues compound it:

• CodeBlock isn't memoized, so any parent markdown re-render re-tokenizes unchanged code.
• ChatWindow rebuilds the toolResults map inline on every render, giving it a new identity each time.

Change

• Code fences render as plain monospace text (same block chrome) while the owning message is streaming; Prism runs once when the message completes. The mermaid source view follows the same rule.
• Wrap CodeBlock in memo.
• Hoist the toolResults map into a useMemo keyed on messages so its identity is stable across streaming updates.

Testing

• tsc --noEmit and eslint clean
• Component tests and lib/markdown.test.mjs pass; added two CodeBlock tests (plain output while streaming, tokenized output once complete)
• Verified on a self-hosted instance (iPhone Safari over remote access): sustained CPU during long code-heavy responses drops noticeably; highlighting appears when the message finishes