Support numpy >= 1.26.4 and Understanding Pinned Dependencies in General #5299

Describe the feature you'd like

Right now, the sagemaker-python-sdk pins certain dependencies to extremely old and restrictive versions. The biggest offender is numpy == 1.26.4 which is over 18 months old, making sagemaker incompatible with many existing codebases.

How would this feature be used? Please describe.

Minimally, would be nice to unpin numpy (permitting latest versions), and in general, understanding why current sagemaker dependencies are pinned to the versions they are currently (e.g., protobuf).

Describe alternatives you've considered

We're considering building our own version of the Sagemaker wheel (with unpinned dependencies), but not sure which of these dependencies will actually break/result in new behavior.

Any help with this would be much appreciated! Thanks!