fix(transcription): strip ***bold italic*** markdown emphasis

[darrenori]

Problem

filter_markdown strips markdown before text reaches TTS, but it only covers *italic* and **bold**. The combined ***text*** / ___text___ form is left untouched, so the delimiters leak through and get spoken.

"This is ***very important*** text."  ->  "This is ***very important*** text."
"___Totally___ critical."             ->  "___Totally___ critical."
"Use ***both*** and **bold**"         ->  "Use ***both*** and bold"

Neither existing pattern can match it:

• bold — the body is [^*\n]+?, which stops at the third asterisk
• italic — the (?!\s|\*) guard rejects a * immediately after the opening delimiter

Same for the underscore variants. LLMs emit ***…*** fairly often for emphasis, so this shows up in practice.

Fix

Add a triple-delimiter pattern ahead of each existing pair. The word/asterisk boundary assertions ((?<![\w*]) / (?![\w*])) are carried over unchanged from the existing patterns.

Tests

Added test_bold_italic, covering the four cases above across chunk sizes 1, 2, 3, 7, 50 so the streaming path is exercised too. All fail before the change, pass after.

Existing coverage is unaffected — the full MARKDOWN_INPUT document test, the punctuation-boundary cases, and the no-false-positive cases (2 * 3 = 6, Use *.py files, x**2 + y**2 = z**2, __dunder_method__) all still pass.

ruff check and ruff format are clean.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.

[longcw]

thanks for the pr! could you sign the CLA before we merge