fix(skills): forward Kokoro TTS speed

[miguel-heygen]

What

Kokoro synthesis now receives the resolved TTS speed at the final subprocess boundary. A requested --speed 1.16 reaches npx hyperframes tts as --speed 1.16 instead of silently using Kokoro’s 1.0 default.

Why

The faceless adapter and shared audio request preserved speed correctly, but Kokoro argv construction dropped it. Synthesis still exited successfully, leaving narration audio and word timing at the wrong pace.

Related: #2335
Related: #2407

How

One Kokoro argument builder owns voice, output, speed, and optional language ordering. synthesizeOne passes its already-resolved speed into that builder, and the default 1.0 value is emitted explicitly under a tested contract.

Test plan

• Exact non-default argv includes --speed 1.16.
• Default argv explicitly includes --speed 1.
• Full provider helper suite passes (13/13).
• Node syntax, oxlint, and oxfmt checks pass.
• Live Kokoro model synthesis (not required; subprocess boundary is deterministic).
• Documentation updated (not applicable).

[miga-heygen]

The fix itself is correct: buildKokoroTtsArgs now always emits --speed, and synthesizeOne forwards its already-resolved speed into it. Tests cover both the non-default (1.16) and explicit-default (1) cases correctly.

However, CI is red: Skills: manifest in sync fails because skills-manifest.json wasn't regenerated after editing skills/media-use/audio/scripts/lib/tts.mjs (hash mismatch for the media-use skill: 00a2d26e22fc1ab1 -> ae2be19b68fe6e84). Please run bun run --cwd packages/cli gen:skills-manifest and commit the updated skills-manifest.json, then this is good to merge.

— Miga