[tpn]

Add CUB Transpose Load/Store so Rust callers can use coalesced block memory access with compiler-managed shared storage:

coop::load(block, input, &mut values,
    coop::LoadOptions::default().algorithm(coop::Transpose));
coop::store(block, output, &mut values,
    coop::StoreOptions::default().valid_items(count).algorithm(coop::Transpose));

Transpose Store requires a mutable payload and permutes both values and validity bits, including slots outside a partial write. Direct remains the default. The standalone example runs both algorithms, and the guide explains scratch reuse and completion.

Review after #12099; this cumulative draft adds typed Transpose recipes and their public API/tests. Requires the matching Oxide toolchain from NVIDIA/cuda-rust#1574. This slice supports complete one-dimensional blocks; other algorithms, warp groups and public storage controls follow separately.

Validated with host tests/doctests, strict lint/docs checks, 20 compiler contrasts, 1,137 Transpose GPU cases and all three CUDA sanitizers on CUDA 13.3/sm_120. The standalone matching-toolchain example validates 3,812 values. Final artifacts retain required barriers and representative kernels have no local spills; measured overhead against numeric CUB remains under investigation.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.