wait_until_completed is not working for metal device

[antimora]

trafficstars

During our benchmark testing, we noticed that the Candle backend for Burn was finishing up quickly for the Metal device. Upon closer inspection, we have discovered that wait_until_completed is not functioning properly. As a fix for now, we have disabled our benchmark for the Candle Metal device (see https://github.com/tracel-ai/burn/pull/1579).

[tomsanbear]

I'm actually in the middle of an overhaul of the metal backend on this PR, could you try out this branch with your benchmark and see if it's still an issue? Note that I've replaced the "wait_until_completed" method with a "synchronize" method to match the cuda backend

https://github.com/huggingface/candle/pull/2037

[antimora]

CC @louisfd

[LaurentMazare]

Note that I'm also adding a synchronize method across devices in #2055, this will ensure that there is a unified and simple way to make these calls.