Add Cerebras fast rewording provider

[MyButtermilk]

Summary

This PR implements the practical path suggested by MyButtermilk in #171: add a very fast built-in provider option for post-dictation cleanup/rewording instead of immediately expanding every prompt with its own provider/model routing.

Changes:

• Add Cerebras as a built-in chat-only provider preset.
• Default Cerebras rewording/chat requests to gemma-4-31b.
• Expose a small curated Cerebras chat model list: gemma-4-31b, gpt-oss-120b, and zai-glm-4.7.
• Keep Cerebras out of transcription flows by marking it CHAT_ONLY.
• Reuse the existing OpenAI-compatible client via https://api.cerebras.ai/v1/.
• Add a focused registry test to prevent the preset from regressing.

Refs #171.

Why this is valuable

Issue #171 discusses per-prompt provider/model selection for rewording prompts. MyButtermilk's comment points out that the largest immediate UX win may not require full per-prompt routing: common post-dictation cleanup tasks are short, repetitive, and latency-sensitive. Punctuation cleanup, filler-word removal, and light grammar correction benefit most from a provider/model path that starts quickly and returns quickly.

Cerebras is a good fit for that narrow path because its inference API is OpenAI-compatible, so Dictate can reuse the existing provider infrastructure instead of adding new request plumbing. That keeps this change small while still giving users a fast rewording target. The selected default, gemma-4-31b, follows the model suggested in the issue discussion and is listed in Cerebras' model catalog.

This improves the app in a few concrete ways:

• Users get a dedicated low-latency cleanup provider without configuring a custom endpoint by hand.
• The existing provider selection UI can surface Cerebras automatically through the registry.
• Transcription provider lists stay correct because Cerebras is registered as chat-only.
• The implementation avoids prompt/profile migrations and avoids increasing the complexity of the prompt editor.
• Future per-prompt routing can still build on this provider preset if the larger feature is implemented later.

Intentional scope

This PR intentionally does not implement full per-prompt provider/model selection. It also does not add provider-specific temperature, max-token, or streaming controls yet. Cerebras supports streaming, but Dictate's current provider preset model is focused on provider capabilities and model defaults, so adding streaming/parameter policy would be a separate UI/configuration change.

Keeping this PR focused makes it easier to review and gives users the main performance-oriented option requested in the issue discussion without expanding the data model.

Validation

Ran:

./gradlew.bat :app:testDebugUnitTest --tests "dev.patrickgold.florisboard.dictate.ProviderRegistryTest" :app:compileDebugKotlin :lib:dictate-core:compileDebugKotlin :wear:compileDebugKotlin

Result: build successful, including ProviderRegistryTest > cerebrasIsAvailableAsFastRewordingProvider().

Also ran git diff --check successfully.

References

• Issue discussion: Allow per-prompt model/provider selection for rewording prompts #171
• Cerebras OpenAI compatibility docs: https://inference-docs.cerebras.ai/resources/openai
• Cerebras model catalog: https://inference-docs.cerebras.ai/models/overview
• Cerebras API keys docs: https://inference-docs.cerebras.ai/console/api-keys
• Cerebras streaming docs: https://inference-docs.cerebras.ai/capabilities/streaming