Fix the norm of complex numbers in clipping

[wdphy16]

Fixes #144

[PhilipVinc]

Bump. can this be merged?

[PhilipVinc]

Bump @mkunesch @rosshemsley , can this be reviewed and eventually merged?

[wdphy16]

Bump @mkunesch @rosshemsley @fabioviola , any update on this?

[PhilipVinc]

Bump again. I'm going to try my luck with @hbq1 .
I'm sorry we are so insistent but we'd love this merged.
Or if optax is more of an internal deepmind project that is not open for external contributions we'd love to know...

[rosshemsley]

Thanks for the pull request, and apologies for the delay in responding.

We've been trying to decide what to do on this issue - one of the challenges is that complex numbers are not yet officially supported in optax, and so we want to be careful of adding support in an inconsistent way to the library.

Optax is also used by a number of models where performance is an important consideration, so there is some concern that this alternative factoring could negatively affect performance of those models in a way that would be hard to test for regressions.

There is wider interest for complex numbers in optax though - (#196), and we're keen to engage with this effort!

One possible (temporary) solution in this case would be to implement an independent clipper specifically for complex numbers. Could that be an option in this case?

[wdphy16]

Closed in favor of #279