fix: send speech keyterms as repeated fields

[Sanjays2402]

Closes #819

The v2.59 client JSON-encodes keyterms into one multipart field, causing the API to apply its per-term length limit to the entire array. The hand-maintained speech-to-text wrapper now routes keyterms through repeated multipart fields without modifying Fern-generated files, for both sync and async clients.

Regression tests inspect the multipart request body for both clients.

---

Note

Low Risk
Localized request-shaping fix in the custom STT wrapper with tests; no auth or core infrastructure changes.

Overview
Fixes speech-to-text keyterms encoding so the API receives one multipart field per term instead of a single JSON-encoded array (which incorrectly applied per-term length limits to the whole list).

The hand-maintained speech_to_text_custom wrapper adds _prepare_keyterms, which strips keyterms from the generated convert call and injects them via request_options.additional_body_parameters. Sync and async convert overrides both use this path without touching Fern-generated code.

New httpx mock tests assert the outbound body has two name="keyterms" parts and does not contain a JSON array string.

^{Reviewed by Cursor Bugbot for commit c877f1e. Bugbot is set up for automated code reviews on this repo. Configure here.}