[Bug] Fix memory OOM in Genesis push_differentiable example

[hughperkins]

Issue: #

Brief Summary

• calling build_stmts twice, on a Call node args, adds an extra 8GB of memory usage when calling push_differentiable.py, https://github.com/Genesis-Embodied-AI/Genesis/blob/main/examples/differentiable_push.py , from Genesis
• we address this by only calling build_stmt on new args, which haven't yet been built, i.e. which don't yet have a ptr attribute
• note that the additional calls to build_stmts were originally added in [Type] Add ndarray struct #73

Before Genesis migrated to gstaichi, in Genesis-Embodied-AI/Genesis#1550, push_differentiable used to use ~22GB of memory, which we can observe by running while true; do { nvidia-smi | grep 575W; sleep 1; } done in parallel, in a second terminal session:

After migrating to gstaichi, this test used ~32GB of memory:

This PR fixes this regression, so this test uses ~22GB of memory again.

The extra memory was being used because we were parsing incoming kernel parameters multiple times. This PR modifies the code so that each kernel parameter is only processed once.

copilot:summary

Walkthrough

copilot:walkthrough

[hughperkins]

We will likely replace this with something like #229

=> draft for now.

[hughperkins]

Actually, it's not clear to me that non-static ranges should fail. So let's merge this anyway. (albeit likely with @yun-long 's suggestion incorporated to make this a bit more principled).

[yun-long]

LGTM

[hughperkins]

Thanks!

[hughperkins]

whoopa, this code has a bug. Not sure how unit tests passed 🤔 Anyway, I will add create a new PR with a unit test that exercises the bug, and a fix for the bug.

[hughperkins]

https://linear.app/genesis-ai-company/issue/CMP-154/fix-bug-with-extend-in-httpsgithubcomgenesis-embodied

[hughperkins]

(In case anyone's curious what the bug is, there shouldn't be an extend here:

_added_kwargs, _kwargs_new = extend(CallTransformer._expand_Call_dataclass_kwargs([kwarg_node]))

)

[hughperkins]

#235