feat(media): add S3 location support to image, document, and video content types

[github-actions]

Motivation

Users need the ability to reference media files stored in Amazon S3 buckets instead of passing raw bytes for image, document, and video content types. This enables scenarios where large media files are already stored in S3 and eliminates the need to transfer bytes directly in API calls.

AWS Bedrock API supports S3 locations for these content types:

• VideoSource
• ImageBlock
• DocumentBlock

Resolves #10

Public API Changes

ImageSource, DocumentSource, and VideoSource TypedDicts now support an optional s3Location field as an alternative to bytes:

# Before: only bytes supported
content = {
    "image": {
        "format": "png",
        "source": {"bytes": image_bytes}
    }
}

# After: S3 location also supported (Bedrock only)
content = {
    "image": {
        "format": "png",
        "source": {
            "s3Location": {
                "uri": "s3://my-bucket/images/photo.png",
                "bucketOwner": "123456789012"  # optional
            }
        }
    }
}

New S3Location TypedDict added:

class S3Location(TypedDict, total=False):
    uri: str          # Required - pattern: s3://[bucket]/[key]
    bucketOwner: str  # Optional - 12-digit AWS account ID

Use Cases

• Large media files: Reference videos or high-resolution images stored in S3 without transferring bytes
• Cross-account access: Use bucketOwner to reference media in other AWS accounts
• Existing S3 workflows: Integrate with pipelines that already store assets in S3

Provider Behavior

Bedrock: Full support - S3 locations are formatted and passed to the API.

Other providers (Anthropic, OpenAI, Gemini, etc.): S3 sources are not supported. When detected, a warning is logged and the content block is skipped:

WARNING - S3 sources are not supported by {provider}, skipping content block