Numpy version discrepancy in Client and Server causes server crash

[hello-fazil]

Uograding the client side numpy version accidentally to 2.2 causes an incompatibility in object serialization/deserialization between client and server. After downgrading the client numpy version to 1.26.4, no exception from the server is observed any more.

This issue needs to be fixed by catching exception and reporting the discrepancy in numpy versions.

[hello-binit]

I think we should be able to use numpy v2, but we'll need to publish a new docker image for the Stretch AI server. Inside the image, there's a __pycache__ that's likely generated w/ np v1, which is the cause of the serialization/deserialization incompatibility.

[hello-peiqi]

Just checked this issue. Looks like few other models such as qwen2.5 conflict with the use of Numpyv2 so I don't think simply removing a __pycache__ would solve the problem.