fix: align wire format for GeminiModel and GeminiCodeAssistModel.

[StarsExpress]

Fixes issue #461 to consolidate duplicated Gemini wire-format logic between gemini_model.py::GeminiModel and gemini_code_assist.py::GeminiCodeAssistModel.

---

Changes

Per issue #461's Proposed Extraction, several functions moved into a new code_puppy/gemini_common.py:

1. _sanitize_schema_for_gemini + _flatten_union_to_object_gemini + generate_tool_call_id

• Previously only in gemini_model.py. Now they are all in gemini_common.py to get called by methods in either gemini_model.py or gemini_common.py, since gemini_model.py depends on gemini_common.py.

• Relationship: _build_tools calls _sanitize_schema_for_gemini, which calls _flatten_union_to_object_gemini.

2. _build_tools

• Shared by GeminiModel and GeminiCodeAssistModel: both emit sanitized schemas under parameters key.

3. _build_generation_config

• Shared by GeminiModel and GeminiCodeAssistModel: reads ModelSettings correctly via .get(), as ModelSettings is a TypedDict, not an object with attributes.

• Additionally handles thinking_level & thinkingConfig.

4. _parse_candidate_parts

• Shared by GeminiModel and GeminiCodeAssistModel: per-part text/functionCall/thought loop + usageMetadata extraction, per issue's exact scope.

---

Tests

New tests/test_gemini_common_full_coverage.py covers shared functions directly, like:

• Schema sanitization.
• Generation config.
• Candidate parsing — thinking parts, function calls with/without API-supplied ids, empty candidates, usage extraction.

Existing test_gemini_code_assist.py also updated where it asserted the old buggy behavior from GeminiCodeAssistModel's drifted usages.

For test_gemini_model_full_coverage.py, since _sanitize_schema_for_gemini + _flatten_union_to_object_gemini + generate_tool_call_id all joined gemini_common.py, their corresponding tests have moved into test_gemini_common_full_coverage.py.