Add configuration to target CUDA compute capability 8.6

[keichi]

This PR adds a flag to compile Kokkos with CUDA compute capability 8.6 (changes are based on #3122). I need this because I'm working with Ampere-based RTX cards.

[dalg24-jenkins]

Can one of the admins verify this patch?

[masterleinad]

Looks good to me.

[dalg24]

OK to test

[keichi]

I ran the unit tests on RTX3090 and the following tests in KokkosCore_UnitTest_Cuda2 failed.

kokkos/core/unit_test/TestTeamTeamSize.hpp

Lines 163 to 167 in 78538c8

	test_team_policy_max_recommended<double, 2, policy_type_1024_2>(0);
	test_team_policy_max_recommended<double, 2, policy_type_1024_2>(
	max_scratch_size / 3 / 2);
	test_team_policy_max_recommended<double, 2, policy_type_1024_2>(
	max_scratch_size / 2);

kokkos/core/unit_test/TestTeamTeamSize.hpp

Lines 178 to 182 in 78538c8

	test_team_policy_max_recommended<double, 16, policy_type_1024_2>(0);
	test_team_policy_max_recommended<double, 16, policy_type_1024_2>(
	max_scratch_size / 3 / 2);
	test_team_policy_max_recommended<double, 16, policy_type_1024_2>(
	max_scratch_size / 2);

After some debugging I figured out cuda_get_max_block_size() was returning zero. I think the reason is that properties.MaxThreadsPerBlock is 1536 on CC 8.6 while it's 2048 on other devices (I've tested on P100 and V100). I modified cuda_deduce_block_size() to account for that but I'm unsure if the fix is correct. Please review.

Changes were added

[dalg24]

LGTM