Infer SageMaker AMIs for CUDA 13 custom images

[shchur]

Summary

• infer SageMaker host AMI versions for custom image tags containing a hyphen-delimited cu13* component
• use al2023-ami-sagemaker-inference-gpu-4-1 for real-time GPU endpoints
• inject al2-ami-sagemaker-batch-gpu-535 for Batch Transform on G4dn, G5, and G6
• preserve explicit user overrides and leave CPU, local, and unrecognized images unchanged
• raise SDK dependency floors to versions that expose the required API fields

This keeps the public workflow unchanged:

predictor.fit(custom_image_uri=image_uri, ...)
predictor.deploy(custom_image_uri=image_uri, ...)
predictor.predict(custom_image_uri=image_uri, ...)

Validation

• uv run --no-project pytest -q tests/unittests/general/test_sagemaker_ami.py tests/unittests/general/test_inference_modes.py tests/unittests/general/test_foundation_model.py
• uv run --no-project ruff check src tests/unittests/general/test_sagemaker_ami.py tests/unittests/general/test_inference_modes.py
• uv run --no-project ruff format --check src tests/unittests/general/test_sagemaker_ami.py tests/unittests/general/test_inference_modes.py
• uv pip check

The billable SageMaker integration suite has not been run as part of this PR yet. The PR is left in draft for validation against the pre-release AutoGluon 1.6 images.

[github-actions]

Job PR-270-4a23111 is done.
Docs are uploaded to https://d12sc05jpx1wj5.cloudfront.net/PR-270/4a23111/index.html