Add custom provider_type for generic OpenAI-compatible endpoints

Problem

Provider catalog currently hard-codes known vendors (Anthropic, OpenAI, DashScope, Claude CLI, ACP, Codex). Users running self-hosted or niche OpenAI-compatible gateways (vLLM, LM Studio, LiteLLM,
Ollama proxy, OpenRouter, Together, Groq, Fireworks, local inference) cannot register them without either (a) hijacking the openai type and mismatching defaults, or (b) shipping a new provider_type per
vendor. Neither scales.

Goal

Introduce a neutral custom provider_type that treats any endpoint as OpenAI chat-completions compatible, so ops teams can onboard new backends via config/UI without code changes.

Scope

In scope

• Backend: accept provider_type = "custom" in provider store + HTTP validation.
• Routing: resolve custom through the existing OpenAI-compatible adapter (HTTP + SSE).
• Model listing: expose /v1/models passthrough for custom endpoints.
• UI: add "Custom Provider" entry in provider dropdown with i18n (en/vi/zh).
• Tests: unit + integration contract coverage.

Out of scope

• Non-OpenAI wire formats (Cohere, Bedrock native, Vertex native).
• Per-custom-provider feature flags (vision, tool-use capability negotiation) — deferred.
• UI credential templates per vendor.

Acceptance Criteria

• Creating a provider with provider_type=custom, base_url, encrypted api_key persists and appears in list endpoint.
• Chat completion request routed to custom provider returns streamed SSE response identical in shape to openai type.
• Model listing endpoint returns models reported by the custom backend.
• UI dropdown shows "Custom Provider" in en/vi/zh; selecting it saves provider_type=custom.
• go test ./internal/http/... ./internal/store/... + go test -tags integration ./tests/contracts/http_api/... pass.
• cd ui/web && pnpm test passes.

Start with the provider store and HTTP validation, then trace how the existing OpenAI-compatible adapter handles HTTP, SSE, and model listing. Review the provider dropdown and its en/vi/zh translations under ui/web. Done means custom providers persist and list correctly, route chat and model requests, appear in all three locales, and pass the specified Go integration and pnpm tests.