fix(py): Fix code execution serialization issue for googleai plugin

[huangjeff5]

Fixes the bug bash entry where ai.generate(config=GeminiConfigSchema(...).model_dump()) with code_execution: True raised:

ValidationError for GenerateContentConfig
codeExecution: Extra inputs are not permitted [type=extra_forbidden]

Root cause: when a request reaches the Gemini plugin, request.config is typed as the generic GenerationCommonConfig, which stores plugin-specific keys verbatim on model_extra (typically alias-form / camelCase). The plugin's _normalize_config_to_dict had a branch that just trusted the base type and dumped it without re-routing through GeminiConfigSchema, so the camelCase extra leaked past the snake_case pop() in _extract_tools_from_config and ended up at the SDK's strict GenerateContentConfig (extra='forbid').

Fix: always re-validate non-plugin inputs through the plugin schema before dumping. Pydantic's populate_by_name=True on the plugin schema folds aliased extras onto their canonical snake_case fields. As a side benefit, this fixes the same latent issue for every other aliased Gemini-only field (googleSearchRetrieval, urlContext, fileSearch, functionCallingConfig, safetySettings, apiKey, contextCache, etc.).

Refactor: extracted _pick_plugin_schema so the routing-by-marker-key logic isn't inlined; the main function now reads as a single top-to-bottom flow.

Tests: 4 parametrized normalize-canonicalization cases (all 3 input shapes + a fast-path GeminiConfigSchema instance) and 1 e2e test that runs _genkit_to_googleai_cfg on the original repro shape and asserts the convenience flag was translated into a Tool(code_execution=...) with no leftover camelCase keys on the SDK config.

Branch: main (apply on top of origin/main).
Preflight: ruff format/check, ty/pyrefly/pyright, pytest 3.10-3.14 -- all green.

[gemini-code-assist]

Code Review

This pull request refactors configuration normalization in the Gemini plugin to ensure camelCase aliases are correctly mapped to canonical snake_case fields. It introduces a _pick_plugin_schema method to select the appropriate schema based on input keys. Review feedback identifies a potential regression for Gemma models because the logic defaults to GeminiConfigSchema, which has stricter validation than GemmaConfigSchema; the reviewer suggests dynamically selecting the schema based on the model version.