feat: expand Vertex/Gemini support — thinking, structured output, tools, caching (0.16.0)

[nyo16]

Summary

Bundle of Gemini-on-Vertex improvements. Most of the new surface lands as helpers in Nous.Messages.Gemini wired into both Nous.Providers.VertexAI and Nous.Providers.Gemini, so anything new works
against either entry point.

Added

• Thinking config (request-side). :thinking_config → generationConfig.thinkingConfig. Accepts Elixir shape (%{thinking_budget: 1024, include_thoughts: true}) or native Vertex shape.
• thoughtSignature round-trip on tool calls. Required for multi-turn thinking + tool loops on Gemini 2.5/3.x. Preserved in tool_call["metadata"]["thought_signature"] and echoed back when serializing
  assistant turns. Stream normalizer also propagates it on :tool_call_delta.
• Structured output. :json_response / :json_schema → responseMimeType / responseSchema. Cross-provider :response_format shapes (%{type: :json_schema, schema: ...} and %{type: :json_object}) map through.
• Safety settings. :safety_settings → top-level safetySettings.
• Tool config / tool choice. :tool_config (raw map) or :tool_choice (:auto / :any / :required / :none / {:any, [\"name\"]}) → top-level toolConfig.
• Function calling actually works on Vertex/Gemini. New Nous.ToolSchema.to_gemini/1 emits proper functionDeclarations shape (strips OpenAI's strict and unsupported additionalProperties).
  Previously the high-level Nous.LLM path silently dropped tools for these providers.
• Native Vertex tools. :native_tools accepts :google_search, :url_context, :code_execution, plus {tool, config} tuples and raw maps. Combined with function declarations in the same tools
  array.
• Context caching. :cached_content → top-level cachedContent (pass-through).
• Streaming + tools. Nous.LLM.stream_text/3 now honors :tools. Tool-call deltas (with thoughtSignature) aggregate per turn, tools execute between turns, conversation continues until the model
  stops calling tools or hits @max_tool_iterations. Text deltas are still yielded as produced.
• More generationConfig fields: topK, seed, candidateCount, presencePenalty, frequencyPenalty, responseModalities.

Changed

• Single timeout source of truth. Removed @streaming_timeout (300s on Vertex, 120s on Gemini). Streaming and non-streaming share one provider default; the actual timeout is model.receive_timeout,
  flowing through build_provider_opts/1.

Version

• mix.exs: 0.15.8 → 0.16.0
• CHANGELOG entry under [0.16.0] - 2026-05-10.

Test plan

• mix test — 1709 tests, 0 failures (101 excluded :llm tag)
• mix compile --warnings-as-errors clean
• mix format --check-formatted clean
• Manual smoke against real Vertex w/ service account: thinking + tool round-trip on gemini-2.5-pro
• Manual smoke: JSON schema response on gemini-2.0-flash
• Manual smoke: :google_search returns grounded citations
• Manual smoke: streaming with tools fires both, completes loop
• Manual smoke: receive_timeout: 600_000 on a long thinking call

Notes

• :tools for Vertex/Gemini was effectively broken in Nous.LLM before this PR (the high-level path used the OpenAI tool schema and the providers never wrote a tools field). After this PR, tool calling
  on Gemini-on-Vertex works through the same Nous.LLM.generate_text/3 and stream_text/3 API as the other providers.
• Helpers added on Nous.Messages.Gemini are public so users can consume them directly if they're calling Nous.Providers.{VertexAI,Gemini}.chat/2 at the low level.