Patch attached in the email.

In some cases, the submit packet can result in configurations changes of 
the hardware decoders. The previous HW context is then freed and a new 
one created. That context is supposed to move up to the main context and 
the other threads.

It appears that when more than 2 frame threads are involved, the new 
context doesn't move in the right place (or rather at the right time). 
And it can create a crash when reusing the old HW context. This patch 
fixes the issue I could reproduce in VLC with DXVA decoding.

I have no idea if this is correct or the side effects induced by this. 
It seems the right thing to do. Keeping the previous call exhibits the 
issue. It seems odd the other existing thread context are not updated 
with the current hwaccel_priv_data. But maybe they are updated later 
from the "main" thread context, in which case this patch seems solid.