fix(py/plugins/google-genai): type the speech config, trim dead Lyria helpers

[hilariie]

Fixes #6185

Stacked on #6353.

Summary

SpeechConfigSchema had no model_config at all, so beyond the two camelCase fields in the issue it also dropped every snake_case key, including the form py/samples/google-genai-media sends for its voice. It now types voiceConfig,
multiSpeakerVoiceConfig and languageCode, matching the SDK's SpeechConfig, and the five orphaned exclude=True fields left on it are gone. One behaviour change: an unknown key nested under speechConfig used to be dropped silently and now raises INVALID_ARGUMENT, since it reaches the SDK's extra='forbid' type.

models/lyria.py still carried its own LyriaConfig and LyriaVersion, unused and now shadowing the Interactions ones that #6241 exports. It keeps only is_lyria_model, which routing needs.

The issue's ImagenConfigSchema bullet is covered by #6305 in the base of this stack, which retires Imagen from Python outright.

[gemini-code-assist]

Code Review

This pull request enhances the Gemini and Imagen configuration schemas by introducing detailed Pydantic schemas and JSON schema annotations. Specifically, it adds support for multi-speaker voice configurations and language codes in Gemini TTS, and explicitly defines numerous configuration fields for Imagen (such as aspect ratio, negative prompt, guidance scale, and safety settings). Additionally, the Lyria model module was refactored to only handle model detection, offloading its configuration and version definitions to the interactions module. Comprehensive unit tests were added to validate the new TTS speech configurations, Imagen configuration mapping, and the package's public API surface. I have no feedback to provide as there are no review comments.