perf: fn Rav1dFrameData::{seq,frame}_hdr: Remove .unwrap() and Arc::deref overhead

[kkysen]

I'm not a fan of these methods. They hide the cost of the .unwrap() and Arc derefs, which we really don't want to be doing in any tight loops like there might be in fn decode_b. Two solutions are:

1. Most of the data accessed through things like frame_hdr are small Copyable fields. These we can just Copy out at the beginning and use directly. For larger fields like arrays, such as gmv, we can pay the cost there.

2. Remove the frame_hdr and seq_hdr fields from Rav1dFrameData. If &Rav1dFrameData can be passed all the way down, then I think frame_hdr and seq_hdr can, and then we can do the .unwrap() and Arc deref only once at the beginning, and it very nicely untangles things.

See #748 (comment).

However, this is just for perf, and the solution is not exceedingly simple, so we can save it for later.

[jplatte]

FWIW, there is ongoing work upstream to make Arc::deref zero-cost.

[kkysen]

FWIW, there is ongoing work upstream to make Arc::deref zero-cost.

Oh very neat, that looks promising.

[jplatte]

Related: If Arc overhead is a concern, have you considered using triomphe? It contains an Arc type without a weak refcount (you don't seem to be using weak references), an OffsetArc type that AFAICT is the same thing as what std is transitioning to (which also looks very similar to your RawArc) + ArcBorrow which can also be interesting for improving perf though I haven't found any place in rav1d yet where it would be immediately applicable.