Fixed-size reductions

[cliffburdick]

Add optimized fixed-size segmented reduce support for CUB 3.2+

Implement conditional support for CUB's new fixed-size segmented reduction API
when using CUB version 3.2 or higher. This optimization applies to both generic
reduce and sum operations when:

• Input is a tensor view
• Both input and output tensors are contiguous memory
• All segments have uniform size

The new implementation uses CUB's simplified API that takes a segment size
parameter instead of begin/end iterators, providing better performance for
uniform segment reductions.

Changes:

• Add version checks for CUB_MAJOR_VERSION >= 3 && CUB_MINOR_VERSION >= 2
• Implement fast path using fixed segment size for contiguous tensors
• Fall back to iterator-based API for non-contiguous or non-uniform cases
• Apply optimization to both DeviceSegmentedReduce::Reduce and ::Sum

This maintains backward compatibility while leveraging newer CUB features
when available.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[cliffburdick]

/build