refactor(models): consolidate duplication into shared httpx/observe/spec layers

[lyingbug]

Summary

Consolidates the internal/models/ package (79 → 55 files, ~10.3k → ~9.0k LOC) without changing any public interface, wire format, or runtime behavior. Adding a new LLM provider now takes a ~15-line declarative spec entry instead of copy-pasting a ~200-line per-provider file.

What changed

Shared infrastructure (new)

• internal/models/internal/httpx/ — one DoPOST with exponential-backoff retry replaces 11 near-identical doRequestWithRetry copies across embedding/rerank providers; a StreamingClient for raw chat streams.
• internal/models/internal/observe/ — generic Wrap / WrapStream decorator machinery replaces 5 Langfuse wrappers + 5 LLM-debug wrappers.
• internal/models/internal/modelconfig/ — one FromModel helper factors out the shared types.Model → Config mapping.

Per-sub-package consolidation

• embedding/: openai.go, jina.go, nvidia.go, azure_openai.go, volcengine.go, aliyun.go (6 files, ~1.2k LOC) → one providers.go declarative spec table + runner.go shared HTTP runner. Ollama and WeKnoraCloud keep their own impls (different transports).
• rerank/: openai/aliyun/jina/zhipu/nvidia reranker files (5 files, ~660 LOC) → providers.go + runner.go. WeKnoraCloud keeps its signer-based impl.
• provider/: 25 single-struct files consolidated into providers.go (data table) + urls.go (BaseURL constants) + model_family.go (IsQwen3 / IsDeepSeek / IsLKEAP* detectors). The 4 struct names (OpenAIProvider, AliyunProvider, MiniMaxProvider, ZhipuProvider) that provider_test.go constructs directly are preserved as thin shells.
• chat/remote_api.go (978 lines) split by responsibility into remote.go (struct + constructor + request builder), remote_nonstream.go (Chat + chatWithRawHTTP), remote_stream.go (ChatStream + streamState + SSE handling), and remote_tools.go (tool-call delta processing + final_answer/thinking streaming).
• chat/, embedding/, rerank/, vlm/, asr/: each sub-package's langfuse_wrapper.go + llm_debug*.go collapsed into one observe.go using the shared decorator.

Guardrails verified

All externally-referenced symbols are unchanged:

• Interfaces: chat.Chat, embedding.Embedder/EmbedderPooler, rerank.Reranker, vlm.VLM, asr.ASR.
• Factories: NewChat, NewEmbedder, NewReranker, NewVLM, NewVLMFromLegacyConfig, NewASR.
• Configs + ConfigFromModel in each sub-package.
• DTOs used by handler/application/agent: chat.{Message, ChatOptions, Tool, ToolCall, FunctionCall, FunctionDef}, rerank.RankResult, asr.{Segment, TranscriptionResult}.
• provider.List(), provider.ListByModelType(), provider.ProviderInfo, all provider.Provider* name constants, provider.WeKnoraCloudBaseURL.
• utils.ChunkSlice (imported by application/service/graph.go and application/service/retriever/...) kept at its current path.
• internal/models/utils/ollama.OllamaService (imported by container/, application/service/model.go, application/service/image_multimodal.go, handler/initialization.go) kept at its current path.
• RemoteAPIChat.BuildChatCompletionRequest (asserted on in remote_api_test.go) kept as a method with the same signature.

Non-goals

• No behavior change. Every existing test in internal/models/ continues to pass.
• No wire-format change for any provider.
• No new features, no removed features, no renamed Config fields.
• No attempt to merge the WeKnoraCloud signer-based transport into the shared runner — its shape differs enough that forcing unification adds complexity.
• No rewrite of chat/ollama.go — the Ollama-API coupling is unavoidable and the file is self-contained.

Test plan

• go build ./... — clean
• go test ./internal/models/... — all pass (chat / embedding / rerank / vlm / asr / provider)
• Diffed against upstream main: the 3 pre-existing test failures in internal/application/{repository,service} (SQLite schema mismatch with wiki_config column) are unchanged — unrelated to this refactor
• Sanity-check a chat completion, embedding, rerank, and VLM call end-to-end against live providers once deployed
• Confirm Langfuse traces and LLM_DEBUG_LOG output shape match pre-refactor format