Implement From<[T; M]> for SmallVec<T, N> for all M, N (v2)

[zachs18]

(First commit removes the unnecessary T: Clone bound from From<array> and From<Vec> impls that weren't there in v1.)

Allow using SmallVec::from(array) and array.into() like you can with Vec, even when array.len() != SmallVec::inline_capacity(). If the source length is longer than the smallvec's inline capacity, it will use heap-allocation, otherwise it will use inline capacity.

const-eval limitations

Ideally, the implementation would be as below to explicitly optimize for the three different cases even in debug mode, but it doesn't work because [T; M] and [T; N] are still considered different types.

if M > N {
  Self::from_vec(Vec::from(array))
} else if M == N {
  Self::from_buf(array) // error: expected `N`, found `M`
} else {
  // same as PR
}

However, with opt-level >= 1, Rust appears to reliably compile this PR to either a basically just memcpy (for M <= N) or an alloc and a memcpy (for M > N).

Backported to v2 as #339

[zachs18]

(force-push explanation: I was going to change this to use the same unconditional with_capacity(m) -> copy_nonoverlapping(..., m) as I did in the v1 backport, but the compiler was not able to reliably optimize it as aggressively as if M > N { Self::from(Vec::from(array)) } else ... at opt-level=1, so I changed it back to that.)

[mbrubeck]

Thanks!