fix(rag): embedding service always forwards `dimensions` even when the model does not support matryoshka

## Bug

In `src/agentscope/app/_service/_embedding.py`, the `build_embedding_model` helper unconditionally sets `kwargs["pass_dimensions"]` based on whether `supported_dimensions` is present on the model card.

This logic is **OpenAI-specific**: only `OpenAIEmbeddingModel` exposes a `pass_dimensions` flag in its `__init__` (see `src/agentscope/embedding/_openai/_model.py:38`). The other three providers (`DashScopeTextEmbedding`, `GeminiTextEmbedding`, `OllamaTextEmbedding`) do **not** accept this kwarg — calling `embedding_cls(**kwargs)` on them raises `TypeError: __init__() got an unexpected keyword argument`.

More fundamentally, when a non-OpenAI provider or a non-matryoshka model is used, forwarding a user-specified `dimensions` can either:

1. **Trigger a provider-side error** if the model has no matryoshka support (BGE, M3E, most OSS embedding models).
2. **Silently produce a different vector size** than the user expects, breaking downstream vector index compatibility.

## Repro

```python
from agentscope.app._service._embedding import build_embedding_model
from agentscope.app.storage import CredentialRecord, EmbeddingModelConfig

credential = CredentialRecord(data={...}, ...)
config = EmbeddingModelConfig(
type="dashscope",
model="text-embedding-v3", # no matryoshka
dimensions=1024, # user-specified
parameters={},
credential_id="...",
)

model = build_embedding_model(credential, config)
# / / With the buggy code: TypeError
# / / dashscope model does not accept `pass_dimensions`
```

## Expected

- Matryoshka-capable models (OpenAI `text-embedding-3-*`, etc.) → forward `dimensions` as configured.
- Non-matryoshka models (BGE, M3E, etc.) → **skip** `dimensions` and use the model default.
- Non-OpenAI providers → no `TypeError`; the service should silently omit the `pass_dimensions` flag for them.

## Proposed Fix

PR #2343 (companion PR):

1. Read `supported_dimensions` from `EmbeddingModelCard` (already exposed in `src/agentscope/embedding/_embedding_model_card.py:66`).
2. Use `inspect.signature(embedding_cls.__init__).parameters` to check whether the embedding class accepts `pass_dimensions`. Only set the kwarg when supported.
3. Set `pass_dimensions=True` only when the model card declares `supported_dimensions`.

This is a 17-line, single-file change (`src/agentscope/app/_service/_embedding.py`) with:

- 100% backward compatibility for OpenAI matryoshka usage.
- Defensive check that future-proofs the service against new providers adding the same flag.
- No behavior change for OpenAI when the model supports matryoshka.

## Environment

- AgentScope 2.0 (current `main` branch, SHA `d3e6bd7`)
- Discovered during 8-week source code deep-dive of the embedding service module.

## Links

- PR: https://github.com/agentscope-ai/agentscope/pull/2343
- Code anchor: `src/agentscope/app/_service/_embedding.py:71-99`

Start in `src/agentscope/app/_service/_embedding.py` at `build_embedding_model` (around lines 71-99) where `kwargs['pass_dimensions']` is currently set and passed to `embedding_cls(**kwargs)`. Then check `src/agentscope/embedding/_embedding_model_card.py` for `supported_dimensions` and confirm constructor support in `src/agentscope/embedding/_openai/_model.py` and other provider constructors used by this service. Re-run the issue's reproduction with a DashScope config to verify no `TypeError`, then test an OpenAI matryoshka case to confirm configured dimensions still flow; done when `pass_dimensions` is only forwarded for supported OpenAI matryoshka models.

由索引模型根據 Issue 內容生成。