Ensure we can optimize out does-nothing `drop`s like `array::IntoIter<NotNeedsDrop, N>`

[scottmcm]

Noticed this looking at #139415.

We already do this for !needs_drop types, but std::array::IntoIter<&mut usize, 1> does have a Drop, and thus needs_drop is true because it's not trivially obvious that the Drop::drop won't actually do anything.

But if you look at https://rust.godbolt.org/z/raWrq6rsY you can see that we're emitting the call to the drop.

This probably is about inlining drop shims in MIR? Might be another way, though, if somehow we could notice in cg_ssa.

[the8472]

Could we help the compiler to make a better needs_drop determination by having an unsafe NoDrop marker trait?

Hrrm, I guess it would be rather limited, a newtype would already defeat that.