[dynamo + optim] complex, sparse are not on tracing testing path

[jon-chuang]

🚀 The feature, motivation and pitch

Even though we allow tracing complex and sparse grads for optimizers, we do not test that these actually work

Solution

Copy some of the tests from test_optim e.g.

https://github.com/pytorch/pytorch/blob/main/test/optim/test_optim.py

into

https://github.com/pytorch/pytorch/blob/main/test/dynamo/test_optimizers.py

Alternatives

Not tested

cc @vincentqb @jbschlosser @albanD @janeyx99 @crcrpar @mruberry @ZainRizvi @ezyang @anijain2305 @chauhang @penguinwu @voznesenskym @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @msaroufim @wconstab @bdhirsh @zou3519 @aakhundov

[janeyx99]

I agree testing should be prioritized, otherwise we should not say they work with no evidence.

If they're not expected to work (not sure on sparse progress with torch.compile), we should at least smoothly fall back to eager vs. erroring.

[janeyx99]

It may be useful to start by enabling PYTORCH_TEST_WITH_DYNAMO across our eager optim tests.