fix: redownload weights only on PytorchStreamReader error

[pawel-kotowski]

Purpose

• Improve error handling in vision agents

Proposed Changes

• Refactored paths handling - it was a bit messy and weights were redownloaded to wrong path as described in Grounddino weight downloaded to wrong location on error during init #678
• Error resulting from broken weights file is detected via "PytorchStreamReader" message in exception (because we always get RuntimeError regardless of error type) in BaseVisionAgent class
• Only in this case the weights are redownloaded
• Other errors are reraised, e.g. CUDA related errors which also prevent model init, even if weights file is correct

Issues

• 678

Testing

• Tests with broken weights files and error related to CUDA_VISIBLE_DEVICES

[maciejmajek]

Refactored paths handling - it was a bit messy and weights were redownloaded to wrong path as described in #678

I think this is still a problem. Interrupting on the first start causes error ".... is a directory"

[pawel-kotowski]

@maciejmajek I added one minor naming fix (didn't affect anything). It works on my side, maybe @jmatejcz could check.

[Juliaj]

@pawel-kotowski, we looked at #678 during Wednesday's dev sync meeting. I was added as a reviewer since I was seeing the CUDA error somewhat frequently when running the RAI manipulation demo a few months ago. Your changes look solid overall, good work! In the PR description, you mentioned still experiencing CUDA errors even with your fixes - is that correct? I haven't been able to reproduce the error with your changes. If you're still seeing the issue, could you run the attached cuda_error_debug.py to help gather diagnostic information?

[Juliaj]

Refactored paths handling - it was a bit messy and weights were redownloaded to wrong path as described in #678

I think this is still a problem. Interrupting on the first start causes error ".... is a directory"

@maciejmajek, how can we reproduce the issue you're experiencing? I tried interrupting the RAIManipulationDemo.GameLauncher startup with Ctrl-C several times but couldn't trigger any error.

[maciejmajek]

Tested again and works! Thank you!