R4-a: gpu-kernels hands every thread a &mut over the whole buffer (aliasing UB)

[dendisuhubdy]

circuits gpu-kernels: all() and dif_tiles give each GPU thread a &mut [u64] over the whole buffer. Writes are disjoint (no data race), but two live &mut to the same memory is undefined behaviour in Rust. Behind the non-default CUDA feature; no node path. Fix: per-element DisjointSlice access, or kernels over raw pointers. Recorded in circuits 27732e9's SAFETY comments.

[dendisuhubdy]

Fixed in code on randprotocol/circuits branch feat/r4a-gpu-aliasing: 7180df8 (a Cells raw-pointer wrapper — every access is a bounds-checked per-element raw-pointer read/write; no thread holds a &mut to the buffer or the shared tile, which is reached through SharedArray::as_raw_mut_ptr) and 02dd38f (the two PTX-build blockers: try_into().unwrap() replaced by explicit arrays in poseidon2_compress/inject, and the Justfile's PTX filename). Verified: host type-check, rand-zkvm-cuda 29/29 (mock-driver) and 25/25 (default), and a compile-only cargo-oxide PTX build for sm_80 of the committed tree, kernel-by-kernel identical to the old code except dif_tiles' two new never-taken bounds checks. Stays open until a GPU run: just kernels sm_80 then cargo test -p rand-zkvm-cuda --features cuda-hw.