Various fixes to run on mps

[jeffra]

Attempting to run DeepSpeed (zero stage 0) with fp32 on MPS device. Ran into a few issues, these fixes resolve them.

• No CUDA-like timer events on MPS, should fall back to host timers
• When abstract accelerator doesn't define a comm backend we shouldn't trigger a broadcast or all-reduce

[tohtana]

Hi @jeffra,
I really appreciate your contribution! This looks good to me, but shouldn’t we show a warning message saying we are skipping weight synchronization and gradient reduction just in case?