fix(api): model-group routing + expose group names in /v1/models

[HeimaoLST]

Two related bugs in the model-group plumbing.

1. keyConfigMiddleware was reading the wrong gin context key

AuthMiddleware (internal/api/server.go:475) sets the authenticated principal under "userApiKey":

c.Set("userApiKey", result.Principal)

But keyConfigMiddleware (internal/api/server.go:1588) was reading "apiKey":

apiKeyRaw, exists := c.Get("apiKey")  // never exists

The middleware therefore early-returned on every request and never populated "apiKeyConfig" / "modelGroup" in the context. Downstream, ginKeyConfigs() in sdk/api/handlers/handlers.go always returned (nil, nil), modelgroup.IsGroupModel was always false, and group names like claude-failover fell through to the normal model lookup which fails with:

unknown provider for model claude-failover  (HTTP 502)

Fix: read "userApiKey" to match what AuthMiddleware sets.

2. Configured model-groups were invisible in /v1/models

The /v1/models endpoint only returned models registered in the global registry. Clients that pick a model from that listing (e.g. Claude Code) had no way to discover that a group name was a valid model identifier.

Fix: add serveModelsWithGroups() helper that appends each configured model-group as a virtual entry to the response:

{
  "id": "claude-failover",
  "object": "model",
  "owned_by": "model-group",
  "type": "model-group",
  "display_name": "claude-failover"
}

The unified models handler now routes through this helper for both the Claude and OpenAI branches so both client families see the groups.

Verification

End-to-end against a local build of this branch with config:

api-key-configs:
  - key: sk-test-key-001
    model-group: claude-failover
    allow-other-models: true
model-groups:
  - name: claude-failover
    models:
      - {model: claude-sonnet-4-6, priority: 3}
      - {model: claude-opus-4-7,   priority: 3}
      - {model: gpt-5.4,           priority: 2}

Endpoint	Result
GET /v1/models	claude-failover, doubao, openai listed alongside real models (type=model-group)
POST /v1/chat/completions {"model": "claude-failover"}	200, routed to claude-sonnet-4-6 (highest priority in the group)
POST /v1/messages {"model": "claude-failover", "stream": true}	200, SSE stream with model=claude-sonnet-4-6 in message_start

Before the fix: every request above returned 502 unknown provider for model claude-failover.

[github-actions]

This pull request targeted main.

The base branch has been automatically changed to dev.