fix: lazy-load text-to-speech websocket peer

[alceops]

Fixes #406.

This keeps the ws peer dependency optional for normal SDK imports by changing the Text-to-Speech websocket wrapper to import ws only as a type and require the runtime package when TextToSpeechWS is actually constructed. If ws is missing at that point, the constructor now throws a targeted optional-peer error instead of making import telnyx fail for every consumer.

I also added a regression test that mocks ws as unavailable and verifies:

• require('telnyx') still succeeds without ws on disk
• constructing TextToSpeechWS is the point where the missing optional peer error is raised

Verification in this environment:

• git diff --check
• static assertions that the top-level value import was removed and the runtime require('ws') is deferred

I could not run the Jest suite locally because this cron runtime does not have Node/npm/pnpm installed.