Implement batched serial/team/teamvector rot

[yasahi-hpc]

This PR implements rot function.

KokkosBatched_Rot.hpp has also been slightly modified to support all of {s,d,cs,zd}rot or {c,z}rot operations.

Following files are added:

1. KokkosBatched_Rot_Impl.hpp: Internal interfaces
2. KokkosBatched_Rot_Internal.hpp: Implementation details
3. KokkosBatched_Rot.hpp: APIs
4. Test_Batched_Rot.hpp: Unit tests for that

Detailed description

It performs the rank 1 operation {s,d,cs,zd}rot or {c,z}rot

x(i) := c*x(i) + s*y(i)
y(i) := c*y(i) - s*x(i) // {s,d,cs,zd}rot
y(i) := c*y(i) - conj(s)*x(i) // {c,z}rot

We check that s is real for {s,d,cs,zd}rot and s is complex for {c,z}rot. c is always real.

Kokkos::parallel_for('rot', 
    Kokkos::RangePolicy<execution_space> policy(0, n),
    [=](const int k) {
      auto sub_x = Kokkos::subview(m_x, k, Kokkos::ALL());
      auto sub_y = Kokkos::subview(m_y, k, Kokkos::ALL());

      KokkosBatched::SerialRot<ArgTrans>::invoke(sub_x, sub_y, c, s);
    });

Tests

1. Make a random x and y, while copying x and y into x_ref and y_ref. The reference x_ref and y_ref are computed by the formula at host. Finally, we confirm x and y computed by serial rot are identical to the reference.
2. Simple and small analytical test, i.e. choose x, and y as follows to confirm x_ref and y_ref are updated as expected.

c = 0.6, s = 0.8
x = [1, 2, 3, 4]
y = [5, 6, 7, 8]
x_ref = [4.6, 6.0, 7.4, 8.8]
y_ref = [2.2, 2.0, 1.8, 1.6]

[lucbv]

I am not quite happy with asking for a Transpose mode when we really only want conj so this needs some re-work

[lucbv]

I do not like calling this ArgTrans, would ArgConj be better since that is actually the only option you have here: regular apply or conjugate apply?
Also it then probably makes sense to default it on NoTranspose?

Suggested change

	template <typename ArgTrans>
	template <typename ArgTrans = KokkosBlas::Trans::NoTranspose>

[yasahi-hpc]

I do not like calling this ArgTrans, would ArgConj be better since that is actually the only option you have here: regular apply or conjugate apply? Also it then probably makes sense to default it on NoTranspose?

That is true.
I just use the boolean to specify the conj operation.

[lucbv]

Suggested change

	static_assert(std::is_same_v<ArgTrans, Trans::Transpose> || std::is_same_v<ArgTrans, Trans::ConjTranspose>,
	static_assert(std::is_same_v<ArgTrans, Trans::NoTranspose> || std::is_same_v<ArgTrans, Trans::ConjTranspose>,

[lucbv]

See above on ArgTrans

[lucbv]

See above regarging ArgTrans

[yasahi-hpc]

Thank you for the review @lucbv
I have updated based on your comments
Can I have another review