fix(anthropic): make Anthropic/Claude provider usable out-of-box (three bugs)

[he-yufeng]

Fixes #2955.

The Anthropic/Claude provider in `lightrag/llm/anthropic.py` had three independent defects that each blocked `anthropic_complete` end-to-end under the current Anthropic SDK and LightRAG's own KG extraction pipeline. All three are addressed here in one file.

1. `voyageai` imported unconditionally at module load

`pipmaster`'s auto-install silently no-ops inside a uv-managed venv (or anywhere without site-packages write permission), so the subsequent `import voyageai` at module top level raised `ModuleNotFoundError` for users who never touched Voyage embeddings at all. Moved the `install + import voyageai` pair into `anthropic_embed` so it's only required when embeddings are actually used.

2. `max_tokens` missing

`anthropic.messages.create()` requires `max_tokens`, but the wrapper never set a default and never surfaced it as a documented kwarg. Every call failed with:

```
TypeError: Missing required arguments; Expected either ('max_tokens', 'messages' and 'model')
or ('max_tokens', 'messages', 'model' and 'stream') arguments to be given
```

Added `kwargs.setdefault("max_tokens", 8192)` — matches the other providers' generous default and callers can still override via `**kwargs`.

3. `stream=True` hardcoded

The wrapper always built `create_params` with `"stream": True` and returned an `AsyncIterator[str]`, so LightRAG's KG extraction (which calls the LLM expecting a `str`) crashed with:

```
TypeError: expected string or bytes-like object, got 'async_generator'
```

A caller that overrode with `stream=False` via `**kwargs` landed on the reverse failure — the SDK returned a `Message` and the code's `async for event in response` blew up with `'async for' requires an object with aiter method, got Message`.

Made streaming opt-in: pop `stream` out of `kwargs`, thread it through explicitly, and return concatenated `content[*].text` when `stream=False` (matching `openai_complete`). The existing async-generator path stays intact for `stream=True`.

Test plan

• Repro from the issue (`rag.ainsert(...)` with `anthropic_complete`) now completes; previously raised at step 2 (`max_tokens` missing) and step 3 (iterator vs string).
• `stream=True` still yields chunks in the pre-existing shape.
• `anthropic_embed` still installs/imports voyageai on first use; LLM-only callers no longer hit `ModuleNotFoundError`.
• `ruff check .` clean on the touched file.

[danielaskdd]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Bravo.

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[danielaskdd]

The dev branch has undergone significant changes related to LLM invocation. Please rebase your current modifications onto dev, conduct thorough debugging and testing, and then submit a pull request targeting the dev branch.