Enforce create_mirror restrictions w.r.t. ViewCtorArgs on all variants

[masterleinad]

In reference to https://kokkosteam.slack.com/archives/G5CBLMFLP/p1689961392307099?thread_ts=1689959585.967619&cid=G5CBLMFLP. We should make sure that we impose the same restrictions on View constructor arguments even if we only return the source View in create_mirror_view. create_mirror_view_and_copy is consistent already.

[fnrizzi]

can you please copy/paste the slack text? so that it is easier to access/read for anybody?

[fnrizzi]

@masterleinad what do mean "even if we only return the source View" ?

[masterleinad]

can you please copy/paste the slack text? so that it is easier to access/read for anybody?

We saw/discussed that

 using MemorySpace = ...;
  auto pif          = 3.14f;
  auto v = Kokkos::create_mirror_view(
      Kokkos::view_alloc("pi", MemorySpace()),
      Kokkos::View<float, Kokkos::HostSpace, Kokkos::MemoryUnmanaged>(&pif));
  ASSERT_TRUE(v.label().empty() || v.label() == "pi");

might compile and run or fail at compile-time depending on MemorySpace.

[masterleinad]

@masterleinad what do mean "even if we only return the source View" ?

If create_mirror_view doesn't create a new object we still want to enforce the same restrictions.

[crtrott]

All of the NVCC builds failed for unrelated reasons, gonna restart this.

[crtrott]

Retest this please

[crtrott]

This time everything passed except the NVHPC run which failed with the OpCode thing from the team algos, and also passed last time around. Also it confirmed again that the OpCode thing seems deterministic on waffle, happened again there.