Summary

Add NVIDIA as a supported provider using their OpenAI-compatible API, with handling for NVIDIA-specific response and request format requirements.

Context

NVIDIA's API endpoint (integrate.api.nvidia.com) follows the OpenAI chat completions format but has two key differences:

1. Response format: NVIDIA sends reasoning_content instead of reasoning, and sometimes sends both fields simultaneously. The previous #[serde(alias)] approach caused deserialization failures when both fields appeared in the same response (serde's "duplicate value" error).
2. Request constraints: NVIDIA rejects requests with multiple system messages or system messages not positioned at the start of the conversation.

Changes

Provider Registration

• Added NVIDIA provider constant with string mappings in ProviderId
• Added NVIDIA entry to provider.json with API URL and NVIDIA_API_KEY environment variable

Response Handling (Serde Fix)

• Replaced #[serde(alias = "reasoning_content")] with a dedicated reasoning_content field on ResponseMessage
• Added merged_reasoning() helper that prefers reasoning over reasoning_content when both are present
• Updated both streaming and non-streaming response paths to use merged_reasoning()

Request Transformer

• Added MergeSystemMessages transformer that consolidates all system messages into a single leading system message
• Wired into the provider pipeline for NVIDIA requests only

Testing

• 4 NVIDIA deserialization tests (tool call streaming, tool call parsing, response message, choice)
• 4 MergeSystemMessages transformer tests (multiple system messages, single system message, no system messages, empty messages)
• 2 provider config tests
• All existing tests updated and passing (551 in forge_app, 282 in forge_domain/forge_repo)

Testing

cargo test -p forge_app -p forge_domain -p forge_repo --lib

All 832 tests pass.