P1-7: Vertex AI / Google AI Studio native generateContent (currently OpenAI-shape relabel only)

Problem

`crates/aisix-provider-gemini/src/lib.rs` is a one-line factory: `OpenAiBridge::new().with_name("gemini")` pointing at `generativelanguage.googleapis.com/v1beta/openai`. The native `:generateContent` API is explicitly out of scope (file header).

Native API exposes:

• `tools` / `function_declarations` (different shape from OpenAI tools)
• `system_instruction` (Gemini-specific)
• `safety_settings` (categorical thresholds — separate from generic guardrails)
• `generation_config` (top_k, response_mime_type, response_schema for structured output)

Customers using Gemini 2.x for structured output / function calling need the native shape.

Scope

DP (moonming/ai-gateway)

• New crate `crates/aisix-provider-gemini-native/` (or extend existing crate) implementing Bridge against `v1beta/models/{model}:generateContent` + `:streamGenerateContent`.
• Two sub-providers: Vertex AI (regional endpoint, OAuth2 service-account) + Google AI Studio (api-key in header).
• Translate OpenAI chat-completion request → Gemini contents/parts shape (and back for response).
• Map system messages → `system_instruction` field.
• Map OpenAI `tools` → Gemini `tools.function_declarations` (recursive JSON schema translation).
• Per-call `generation_config` from request `response_format` field.

CP (api7/api7ee-3-control-plane)

• Provider enum: add `vertex_ai` and `google_ai_studio` (or just `gemini_native` with sub-mode).
• Credential schema: Vertex needs service-account JSON; Studio needs API key.

Dashboard UI (api7/AISIX-Cloud)

• Provider picker adds the native variants.
• Credential form per variant.

e2e

• Tool-calling test against mock Gemini.
• structured-output test (`response_schema`).

Estimate

DP 5-7d, CP 1d, UI 1d, e2e 1d

Start with crates/aisix-provider-gemini/src/lib.rs and the Bridge contract, then trace how provider crates, control-plane enums, dashboard credentials, and e2e tests are organized. Done means native Vertex AI and Google AI Studio generateContent and streamGenerateContent support, request/response mappings, credentials and provider variants, plus mock tool-calling and structured-output tests.