fix(Connectors.Google): use embedContent API for gemini-embedding models - #14266

What changed

VertexAIEmbeddingGenerator hardcodes the :predict endpoint for every model. That method is no longer served for gemini-embedding-* models, so using e.g. gemini-embedding-2 fails with 400 FAILED_PRECONDITION.

This PR routes embedding requests by model family:

• gemini-embedding-* models → the :batchEmbedContents method, with the requests[].content.parts[].text payload shape and the embeddings[].values response shape.
• All other models (text-embedding-*, multimodalembedding, etc.) → unchanged :predict path with the existing instances payload.

Why this shape

batchEmbedContents is the batched equivalent of embedContent — it maps 1:1 onto the client's existing behavior of embedding a list of strings in a single HTTP call, and returns embeddings in input order. outputDimensionality is forwarded exactly as before.

Tests

Added to VertexAIClientEmbeddingsGenerationTests (red → green):

• ShouldUseBatchEmbedContentsEndpointForGeminiEmbeddingModelAsync — asserts :batchEmbedContents endpoint + requests[].content.parts[].text body for gemini-embedding-2.
• ShouldUsePredictEndpointForLegacyEmbeddingModelAsync — asserts :predict endpoint + instances body for text-embedding-004 (regression guard).
• ShouldReturnValidEmbeddingsResponseForGeminiEmbeddingModelAsync — parses a new vertex_embed_content_response.json fixture into vectors.

Full Connectors.Google.UnitTests suite: 437 passed, 0 failed.

Fixes #14265