feat: add Ollama backend (text + vision) with shared transport refactor

[NakaokaRei]

Summary

Add Ollama as a fourth AI backend alongside FoundationModelsBackend, OpenAIBackend, and OpenAIVisionBackend, giving users a free, private, offline alternative for both ActionGenerator and the Agent loop. Ollama exposes an OpenAI-compatible /v1/chat/completions endpoint, so no SDK dependency is required.

Approach: extract shared logic first

Today's two OpenAI backends each contain ~30 lines of duplicated HTTP transport, envelope parsing, and action parsing. Naively adding two more files would 2x that duplication. Instead, this work will:

1. Extract shared HTTP / JSON / parsing logic into Sources/SwiftAutoGUI/Internal/ (no public API change):
   • ChatHTTP — POST + Bearer auth + status check + JSON envelope parse
   • JSONUtilities — robust JSON extraction (handles SSE / trailing garbage)
   • ActionDictParser — action dict -> BasicAction
   • ActionContentParser — inner JSON content -> [Action] / AgentResponse
   • ResponsesAPIDecoder — extract content from OpenAI Responses API envelope
   • ChatCompletionsDecoder — extract content from Chat Completions envelope
   • ActionSchemas — JSON schemas (action item, agent response wrapper)
   • SystemPrompts — actionGeneration and agentLoop(...) prompts
2. Refactor OpenAIBackend and OpenAIVisionBackend to thin wrappers over these helpers (file size drops by ~60%).
3. Add OllamaBackend and OllamaVisionBackend as similarly thin wrappers — provider-specific code is just message building + endpoint URL.

After the refactor, each concrete backend file contains only what is genuinely provider-specific.

Scope

• OllamaBackend conforming to ActionGenerating (text -> action sequence).
• OllamaVisionBackend conforming to VisionActionGenerating (screenshot + history -> agent response).
• Default base URL http://localhost:11434/v1; optional apiKey parameter for proxied/cloud Ollama deployments.
• --backend ollama flag added to the sagui agent CLI.
• Backend selector added to the Sample app (AgentDemoView / AgentDemoViewModel).
• init(ollamaModel:baseURL:apiKey:) convenience init on ActionGenerator.
• Updates to README.md and DocC documentation (Documentation.docc/SwiftAutoGUI.md).
• Unit tests for the new shared helpers and Ollama-specific request building (no network).

Out of scope

• Streaming responses.
• Auto-detection of installed Ollama models.
• Pulling models from the Ollama library.

Verification

• swift build — package compiles after refactor.
• swift test — pre-existing tests pass; new SharedHelpersTests and OllamaBackendTests pass.
• swift package generate-documentation — DocC builds without missing symbols.
• Manual smoke test with a local Ollama install:

  ollama pull llama3.2-vision
  swift run sagui agent "open Finder" --backend ollama --model llama3.2-vision

• Sanity-check the OpenAI path still works after the refactor (regression catch).

Notes

• No Package.swift dependency is added — Ollama's OpenAI-compatible endpoint is consumed via URLSession, matching the existing OpenAI backends.