tts: strip emoji, diagrams, code, tables from TTS preprocessing

[nnnet]

Summary

_strip_markdown_for_tts only handled basic markdown (bold/italic/headers/lists/code-fences/links) but left through everything that TTS engines (edge, kokoro, piper, xtts) read as garbage: emoji, dingbats, box-drawing diagrams, geometric shapes, arrows, math operators, and ASCII divider lines.

Symptoms reported in production:

• "✅ Готово" read as "галочка большая Готово"
• "⚙️ mcp_filesystem_edit_file" read as "медный значок"
• Diagram ┌─┐ │ box │ └─┘ → long awkward pauses with "box" mangled
• Inline code → identifiers pronounced character-by-character

This patch extends the function to also drop:

• inline backtick code (entire snippet, not just backticks)
• markdown tables (entire rows — TTS can't read columnar layout)
• blockquote prefix >
• Unicode pictographs / dingbats / box-drawing / geometric shapes / arrows / math symbols / variation selectors / ZWJ joiners
• pure-divider lines (────── or ============)

The original chat output is untouched — the function runs only on the copy passed to the TTS engine. Telegram/Discord/terminal messages remain visually formatted; only the spoken reply gets clean prose.

Test plan

• Manual tests with Cyrillic + ASCII text containing emoji, headers, lists, code-blocks, tables, diagrams — all stripped correctly
• Run upstream test suite (if any covers _strip_markdown_for_tts)
• Real-world test with edge-tts engine — confirm no regressions on normal text

[alt-glitch]

Duplicate of #8205 which is the most comprehensive TTS preprocessing PR (strips markdown, localizes 28 emojis across 13 languages). Also overlaps with #13311 and #18598. Additionally, this PR contains 8 unrelated file changes (.env.example, sync.yml, auxiliary_client.py, codex.py, gateway/run.py, config.py, dashboard dist, run_agent.py) beyond the TTS scope.