Cache syntax highlighting and markdown rendering per message

[bug-ops]

Parent: #499

Problem

Every frame re-runs tree-sitter syntax highlighting and pulldown_cmark markdown parsing for all visible messages. For responses with multiple code blocks, this can take tens of milliseconds per frame, compounding with the event loop issue (#500) to cause UI freezes.

Locations

• crates/zeph-tui/src/widgets/chat.rs: render_md() called per message per frame
• crates/zeph-tui/src/widgets/chat.rs: push_code_block_text() calls SYNTAX_HIGHLIGHTER.highlight() per code block per frame
• crates/zeph-tui/src/widgets/chat.rs: wrap_spans() iterates char-by-char per line per frame

Fix

Add a render cache to App or ChatMessage:

• Cache key: message content hash + terminal width
• Cache value: pre-rendered Vec<Line<'static>>
• Invalidate when: content changes (streaming), terminal resizes, tool expanded/compact toggles
• Streaming messages: cache only after streaming flag becomes false

This eliminates redundant markdown parsing, syntax highlighting, and word wrapping for finalized messages.

Acceptance Criteria

• Finalized messages are rendered from cache (no re-parsing)
• Streaming messages still update in real-time
• Cache invalidates on terminal resize
• Frame time measurably decreases with many code blocks visible