feat: parallel TTS prefetch + math-to-speech prompt rules

[elimelt]

Summary

Two improvements to the TTS reader, following on from the serial queue +
constrained rewriting work in #143.

1. Smoother playback — parallel prefetch queue

The old queue synthesized one chunk at a time and stayed only 2 chunks
ahead, so audio stalled whenever playback caught up to the single
in-flight LLM+TTS pipeline.

• CONCURRENCY = 4 synthesis pipelines now run in parallel, each
  claiming the next un-started index via a next pointer.
• LOOKAHEAD = 6 keeps a deep buffer ahead of playback.
• Exactly-once is preserved: const i = next++ claims a distinct
  index per worker with no await in between (JS is single-threaded).
• In-order playback is unchanged: awaitChunk(i) still waits for
  index i regardless of the order results land.
• Epoch cancellation, waiter wake-ups, and object-URL cleanup all
  preserved; stop() resets next.

Safe now that the upstream rate limit was raised ~10x. Lower
CONCURRENCY if 429s reappear.

2. Better math-to-speech prompt

LLM_PROMPT now lists explicit conversion rules instead of relying on a
couple of examples:

• Subscripts: x_i -> "x i"
• Powers: x^2 -> "x squared", 2^n -> "two to the n"
• Function application: f(x) -> "f of x", g(x, y) -> "g of x and y"
• Big-O, fractions, relational/arithmetic operators
• Common LaTeX commands (\sqrt, \sum, \approx, \to, ...)
• Greek letters read by name; code identifiers with separators spoken as
  spaces

Plus an anti-refusal directive so short technical fragments aren't
mistaken for instructions. CACHE_VERSION bumped to 3 so rewrites cached
under the old prompt are ignored.

Testing

• Standalone simulation of the parallel queue: exactly-once production,
  concurrency capped at 4 (real parallelism observed), correct under
  variable per-chunk latency, clean stop/restart.
• Live LLM check of the new prompt: all requested rules apply and plain
  prose is left verbatim. Remaining small-model flakiness (occasional
  refusal on a bare a != b) is caught by the existing length-ratio
  guard, which falls back to reading the raw text.
• npm run build clean (1700 files); new logic confirmed in the shipped
  bundle.