Fix block reduce alignment

[gevtushenko]

Description

closes #1229

Vectorized loads of temporary storage in block reduce, lead to unaligned access. This PR marks temporary storage for block reduce as struct alignas(detail::max_alignment_t<16, T, WarpReduceStorage>::value) _TempStorage. This and reordering of member variables in temporary storage, ensures that warp aggregate loads can be vectorized.

Note that other parts of CUB rely on __align__(16). This is potentially problematic for custom types with explicitly specified alignment. I'll create a separate issue for unifying the approach later.

Checklist

• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[gevtushenko]

Performance is equivalent after the change:

## [0] NVIDIA RTX 6000 Ada Generation

|  T{ct}  |  OffsetT{ct}  |  Elements{io}  |   Cmp Noise |   %Diff |  Status  |
|---------|---------------|----------------|-------------|---------|----------|
|   I8    |      I32      |      2^24      |       2.41% |   0.28% |   PASS   |
|   I8    |      I64      |      2^24      |       3.02% |   0.19% |   PASS   |
|   I16   |      I32      |      2^24      |       2.67% |   0.25% |   PASS   |
|   I16   |      I64      |      2^24      |       2.52% |  -0.12% |   PASS   |
|   I32   |      I32      |      2^24      |       2.42% |   0.03% |   PASS   |
|   I32   |      I64      |      2^24      |       2.69% |  -0.05% |   PASS   |
|   I64   |      I32      |      2^24      |       1.68% |   0.00% |   PASS   |
|   I64   |      I64      |      2^24      |       1.77% |   0.01% |   PASS   |
|  I128   |      I32      |      2^24      |       1.81% |   0.05% |   PASS   |
|  I128   |      I64      |      2^24      |       1.63% |  -0.02% |   PASS   |
|   F32   |      I32      |      2^24      |       2.48% |   0.06% |   PASS   |
|   F32   |      I64      |      2^24      |       2.78% |  -0.11% |   PASS   |
|   F64   |      I32      |      2^24      |       1.88% |   0.04% |   PASS   |
|   F64   |      I64      |      2^24      |       2.04% |  -0.06% |   PASS   |
|   C64   |      I32      |      2^24      |       1.88% |  -0.09% |   PASS   |
|   C64   |      I64      |      2^24      |       1.70% |   0.03% |   PASS   |
|   I8    |      I32      |      2^28      |       1.80% |  -0.04% |   PASS   |
|   I8    |      I64      |      2^28      |       2.64% |  -0.22% |   PASS   |
|   I16   |      I32      |      2^28      |       1.91% |  -0.14% |   PASS   |
|   I16   |      I64      |      2^28      |       1.81% |   0.01% |   PASS   |
|   I32   |      I32      |      2^28      |       1.39% |   0.06% |   PASS   |
|   I32   |      I64      |      2^28      |       1.48% |  -0.16% |   PASS   |
|   I64   |      I32      |      2^28      |       9.73% |  -0.94% |   PASS   |
|   I64   |      I64      |      2^28      |       8.91% |   1.64% |   PASS   |
|  I128   |      I32      |      2^28      |       7.61% |   1.08% |   PASS   |
|  I128   |      I64      |      2^28      |       7.08% |   0.02% |   PASS   |
|   F32   |      I32      |      2^28      |       1.07% |   0.12% |   PASS   |
|   F32   |      I64      |      2^28      |       1.38% |   0.06% |   PASS   |
|   F64   |      I32      |      2^28      |       9.92% |  -1.05% |   PASS   |
|   F64   |      I64      |      2^28      |       9.02% |   0.48% |   PASS   |
|   C64   |      I32      |      2^28      |      10.04% |  -0.40% |   PASS   |
|   C64   |      I64      |      2^28      |       9.21% |   0.51% |   PASS   |

[miscco]

Suggestion:

we usually go with the ordering of static constexpr because static is the more relevant information here

[bernhardmgruber]

Remark: clang-format can handle such orderings using the QualifierOrder style option, which we seem to not set in our .clang-format.

[bernhardmgruber]

Opened: #1748

[gevtushenko]

Addressed by nvbug 4428282