v0.8.499

• Atom-JEPA model framework: template, trainer, UQ, registry weights, training image deps, tests and docs
• Training jobs get a 24 hour runtime cap, matching the Batch job ceiling
• Adding a timer to our boto3 inference script
• adding realtime support to the MetaEndpoint class
• Some additional work the AtomJEPA
• Local endpoints put the model dir on the import path, matching the container
• Adding python3-dev to our pytorch-chem training images
• Merge branch 'atomjepa_finetune'
• Grant Glue roles ec2:DeleteNetworkInterface so VPC connection ENIs are cleaned up; add orphaned Glue ENI cleanup script
• AtomJEPA spike work
• AtomJEPA spike work
• Move no-Workbench endpoint examples to examples/standalone_inference; add boto3 example