feat: add support for multimodal model and add embed_image

[svidiella]

Add support for multimodalembedding model as model type on VertexAIEmbeddings.

• Add embed_image method, only valid for multimodalembedding model.
• Make embed_documents and embed_query only available for text models.
• TODO: adapt embded_documents and embded_query to also work with multimodalembedding models.

[holtskinner]

Not sure it makes sense for this to be a static method. This could be an internal @property that just checks self.model_name

[lkuligin]

nits: maybe it's time to add a mapping of model names to different properties in one place. we have too many is_this or is_that conditions all over the place.

[svidiella]

@holtskinner the VertexAI model is happening before the class has been initialized, inside the pydantic validator, so I was assuming I won't have access to the self.model_name at that point.
@lkuligin to be addressed in another PR? What approach were you thinking about?

[lkuligin]

@svidiella another PR for general cleanup - sure. maybe smth similar to this approach:

langchain-google/libs/genai/langchain_google_genai/llms.py

Line 24 in 75d0d8c

class GoogleModelFamily(str, Enum):

for this PR - maybe just to keep it consistent, take a look how it's done for chat models / llms, please.
there is a property that uses is_smth method imported from utils (to address exactly the point you've mentioned).

[svidiella]

@lkuligin I took the approach from GoogleModelFamily, didn't implement the is_smth method at the end
Take a look and tell me if you prefer the other apprach.