Add on-device Whisper dictation for the chat composer

[OurBigAdventure]

Summary

• Adds a Dictate control that fills the chat composer from the microphone using Whisper Tiny in a local Web Worker.
• Speech stays on the device: env.allowRemoteModels is disabled, and weights are staged at build time under /assets/whisper/models/ instead of calling a cloud speech API.
• Uses fp32 ONNX graphs (encoder_model.onnx, decoder_model_merged.onnx). Quantized (q8) Whisper graphs fail under the current Transformers.js / ONNX Runtime combination (MatMulNBits missing scale).

This is separate from Add Voice, which attaches Codec2 or Opus audio to an LXMF message.

Why

Field use needs speech-to-text without sending audio off the device.

Test plan

• Open a conversation, click Dictate, speak, stop, and confirm text lands in the composer.
• Confirm the first use may take a few seconds to load the local model, and later uses reuse it.
• Confirm dictation still works after a frontend rebuild (npm run download-whisper-assets then Vite).
• Confirm Add Voice is unchanged.