feat(voyageai): add multimodal image embedding support

[markritterman]

Summary

The Embeddings API already supports image inputs via fromImage() / fromImages() (added in #789), but the VoyageAI provider was ignoring them and only sending text to the /embeddings endpoint.

This PR updates the VoyageAI embeddings handler to route image-containing requests to Voyage's /multimodalembeddings endpoint with the correct content structure:

• Base64 images — sent as data:{mime};base64,... format
• Image URLs — sent as image_url type
• Mixed text + image — each input becomes a separate multimodal content object
• Text-only requests — unchanged, still use /embeddings
• Provider options (inputType, truncation) forwarded to both endpoints

Example usage

use Prism\Prism\Enums\Provider;
use Prism\Prism\Facades\Prism;
use Prism\Prism\ValueObjects\Media\Image;

// Image embedding
$response = Prism::embeddings()
    ->using(Provider::VoyageAI, 'voyage-multimodal-3')
    ->fromImage(Image::fromStoragePath('photos/sunset.jpg', 's3'))
    ->withProviderOptions(['inputType' => 'document'])
    ->asEmbeddings();

// Text query embedding (for searching against image embeddings)
$response = Prism::embeddings()
    ->using(Provider::VoyageAI, 'voyage-multimodal-3')
    ->fromInput('sunset over the ocean')
    ->withProviderOptions(['inputType' => 'query'])
    ->asEmbeddings();

Test plan

• Image via base64 routes to /multimodalembeddings
• Image via URL routes to /multimodalembeddings with image_url type
• Mixed text + image returns multiple embeddings
• Text-only still uses /embeddings (regression)
• Provider options forwarded to multimodal endpoint
• All 1355 existing tests pass