fix(0.0.7): derive real chain/model name from langchain.js callbacks

[Dank-del]

What

Fix the long-tail of `chain_type: "unknown"` and `model: "ChatOpenAI"` (the wrapper class instead of the configured model id) in AxonPush traces emitted from LangGraph and modern Chat* integrations.

Sister PR to axonpush/python-sdk#6 — same bug, same fix.

Why

LangChain.js's `handleChainStart(chain, inputs, runId, parentRunId, tags, metadata, runType, runName)` was originally designed for plain `Chain` classes that populate `serialized.name`. LangGraph nodes — which compile down to anonymous Runnables — instead pass `serialized={}` and put the node identity into the trailing `runName` arg + `metadata.langgraph_node`. The handler was reading only `serialized.name` and falling back to the literal string `"unknown"`, so every graph step in the dashboard rendered as `chain_type: "unknown"`.

Same problem applied to `handleLLMStart`: `serialized.name` returns the wrapper class (`ChatOpenAI` / `ChatAnthropic` / …), not the model id. The actual model lives in `extraParams.invocation_params.{model,model_name}` at call time.

How

Three helpers in `src/integrations/_base.ts`:

```ts
deriveRunnableName(serialized, runName, metadata)
// runName -> metadata.langgraph_node -> serialized.name
// -> last segment of serialized.id -> "Runnable"

deriveModelName(serialized, extraParams)
// invocation_params.model/model_name -> extraParams.{model,model_name}
// -> serialized.kwargs.{model,model_name} -> serialized.name -> "unknown"

extractRunMetadata(tags, metadata, runType)
// pulls langgraph_{node,step,triggers}, thread_id, run_type, tags
// into the event metadata dict so the UI can group/filter by node + tag
```

`handleChainStart` and `handleLLMStart` (in both `langchain.ts` and `langgraph.ts`) now accept the trailing positional args LangChain.js already passes (`tags`, `metadata`, `runType`, `runName` for chain start; `extraParams`, `tags`, `metadata`, `runName` for LLM start) and use the helpers above. Old callers that ignore those args keep working — purely additive at the public surface.

Test plan

• `bun run test` — 173 pass (165 + 8 new)
• `bun run build` — clean
• New tests in `src/tests/integrations/_base.test.ts` cover:
  • LangGraph empty-`serialized` path with `runName` + `metadata.langgraph_node`
  • `extraParams.invocation_params` model resolution
  • Each fallback rung of both helpers
  • `extractRunMetadata` round-trip
• Manual: re-run a LangGraph agent in `@axonpush/sdk` consumer, confirm dashboard shows real node names and model ids.

Migration

None. Purely additive. Existing callers keep working; new trailing args are all optional.