Attempt Polymorphization

[james7132]

Partially fixes #78.

This attempts to minimize the amount of monomorphized codgen created while using this crate.

• Header<M> has been split into Header and HeaderWithMetadata<M>, using repr(C) to keep pointers trivially castable between each.
• Move any header-only operations onto the Header type.
• Move out multiple entries of the TaskVtable to operate without every generic argument on RawTask, ,just the ones they need.
• Make functions like get_output a function on the vtable itself instead of put in the vtable, both avoids an extra dynamic dispatch call, and shrinks the vtable.
• Embed the RawWakerVtable in the TaskVtable and pass it into polymorphized functions.
• Use #[inline(never)] on these polymorphized functions.

This PR also shifts to use byte_add, which does bump the MSRV to 1.75. If this is too much of a bump for such a gain, I can revert that.

Benchmarks:

master
test task_create ... bench:          35.94 ns/iter (+/- 0.53)
test task_run    ... bench:          40.06 ns/iter (+/- 0.55)

polymorphization
test task_create ... bench:          37.36 ns/iter (+/- 1.69)
test task_run    ... bench:          40.63 ns/iter (+/- 1.14)

[notgull]

Looks good to me overall. Sorry for the delay.

Do we have evidence that this actually reduces polymorphization in practice? Like, if we compile Bevy before and after this patch, how much does the code size decrease.

[notgull]

Mark this as unsafe

[james7132]

Just went ahead and deleted the function. It's was only used in one location.

[james7132]

@notgull there is the (admittedly artificial) benchmark results seen here: #66 (comment).

Bevy actually uses surprisingly few new instances of tasks (generally not used in generic code), and has other compilation bottlenecks.

[james7132]

Went ahead and reverted the byte_add changes and instead chose to polyfill it. I don't think it's worth bumping the MSRV of the entire crate for.

[notgull]

Looks good to me!