inlined_vector: Use trivial relocation for SwapInlinedElements

[Quuxplusone]

I noticed while working on #1615 that inlined_vector could use the trivial relocatability trait here, too.
Here the memcpy codepath already exists; we just have to opt in to using it.

[derekmauro]

Going to try to import this now. I don't expect any issues.

[Quuxplusone]

Awesome, 92c8575 LGTM!
FWIW, re this comment:

// TODO(absl-team): Using unique_ptr here is technically correct, but
// a trivially relocatable struct would be less semantically confusing.
TEST(UniquePtr, Swap)

notice that we definitely don't want to use a trivially copyable struct, because that wouldn't sufficiently test the behavior we're trying to test. I assume the tests in the UniquePtr group use unique_ptr heap-allocation because we can easily valgrind the tests to double-check that each element's value is destroyed once (not twice or zero times). At least I hope it's true that you can easily (and nightly) valgrind the tests. If not, I'd be the first to agree that all this stuff is woefully under-tested! :)

[derekmauro]

FWIW, re this comment:

// TODO(absl-team): Using unique_ptr here is technically correct, but
// a trivially relocatable struct would be less semantically confusing.
TEST(UniquePtr, Swap)

notice that we definitely don't want to use a trivially copyable struct, because that wouldn't sufficiently test the behavior we're trying to test. I assume the tests in the UniquePtr group use unique_ptr heap-allocation because we can easily valgrind the tests to double-check that each element's value is destroyed once (not twice or zero times). At least I hope it's true that you can easily (and nightly) valgrind the tests. If not, I'd be the first to agree that all this stuff is woefully under-tested! :)

A code reviewer suggested that comment, but you are correct. Maybe it would be better to static_assert that the unique_ptr has the properties that we want for testing as a form of self documentation, along with a comment about why we choose that type. We don't have valgrind testing, but we do have MSAN testing, which (I think?) is just as good.