port upstream hecs performance improvements

[cart]

Hecs just introduced some pretty significant performance improvements in:

Ralith/hecs#93 (@mjhostet)
Ralith/hecs#95 (@Ralith)
Ralith/hecs#96 (@Ralith)

[cart]

My big hope is that this will finally allow us to merge QueryBorrow and QueryIter without killing performance, enabling:

// query.iter() returns an actual iterator
for x in query.iter() {}

I'm testing that now

[Ralith]

merge QueryBorrow and QueryIter

FWIW these are distinct for correctness, not performance, reasons. The references yielded by iteration must have lifetimes bound to something more short-lived than the world to prevent illegal aliasing, and they can't be bound to the iterator object directly because Iterator doesn't support that.

[cart]

FWIW these are distinct for correctness, not performance, reasons. The references yielded by iteration must have lifetimes bound to something more short-lived than the world to prevent illegal aliasing, and they can't be bound to the iterator object directly because Iterator doesn't support that.

Thanks for the heads up! We've accounted for the correctness part by removing locks entirely on world apis / relying on lifetime protection for mutable queries. Bevy's higher-level system queries also prevent unsafe iteration.

However merging them (prior to this pr) severely tanked performance for some reason.