[Studio] Add real LLM gateway

[Aias00]

Summary

• add an OpenAI-compatible LLM client for openai/deepseek/tongyi/ollama chat completions
• add a primary LLM gateway that uses the real provider when config is complete and falls back to the stub only when config is incomplete
• preserve the existing SSE contract by emitting message events with {"text": ...} and done events with [DONE]
• add structured LLM failure metadata (code, hint, statusCode) for gateway errors and config-test failures
• map LLM gateway failures through the global exception handler instead of falling back to generic 500 errors
• surface config-test failure hints and machine-readable codes in the LLM settings UI
• validate provider base URLs, token limits, and temperature before saving/testing LLM settings
• normalize pasted OpenAI-compatible /chat/completions endpoints back to the provider base URL before requests are sent
• load model options from the configured OpenAI-compatible provider /models endpoint, with built-in fallback models when the provider is unavailable
• use saved LLM provider/model configuration on the AI chat page instead of hard-coded model choices

Scope

• Track 3 / RIP-3 AI Native
• Covers the first backend gateway slice of [Studio][RIP-3] LLM gateway to replace the stub (multi-provider via spring-ai) #545: OpenAI-compatible providers only
• Providers covered in this slice: openai, deepseek, tongyi, ollama

Failure handling covered

• missing provider API key returns and displays llm.config.missing_api_key with a remediation hint
• invalid provider base URL returns llm.config.invalid_api_base
• out-of-range maxTokens returns llm.config.invalid_max_tokens
• out-of-range temperature returns llm.config.invalid_temperature
• unsupported configured providers return llm.config.unsupported_provider
• upstream HTTP errors preserve upstream status and message where available
• provider timeout returns 504 / llm.provider.timeout
• empty provider response returns llm.provider.empty_response
• non-JSON provider response returns llm.provider.malformed_response
• provider model listing uses GET {apiBase}/models and falls back to built-in provider models on timeout, upstream error, or malformed model responses
• frontend API client passes through domain payloads that use string code fields, while retaining numeric Result envelope handling

Non-goals

• no tool calling in this slice
• no multi-turn history changes
• no Azure/Bedrock native adapters yet

Tests

• Latest AI page wiring: 447883a [Studio] Use LLM config on AI chat page
• Latest backend provider-model enhancement: 8dfea98 [Studio] Load LLM models from provider
• JAVA_HOME=$(/usr/libexec/java_home -v 21) mvn -q -Dtest=OpenAiCompatibleLlmClientTest,LlmConfigServiceTest,OpenAiCompatibleLlmGatewayTest,LlmControllerTest,GlobalExceptionHandlerTest test
• JAVA_HOME=$(/usr/libexec/java_home -v 21) mvn -q test
• npm test -- --run src/api/llm.test.ts src/api/client.test.ts src/pages/studio/__tests__/LlmSettings.test.ts
• npm run build
• npm run lint
• git diff --check

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[Aias00]