ROCm Composable Kernel user guide (CK 1.2.0 documentation)

AMD's Composable Kernel: 'a programming model for writing performance critical kernels for machine learning workloads across multiple architectures including GPUs and CPUs, through general purpose kernel languages such as HIP C++'. That is the honest category statement - CK is a header-only C++ template library in HIP C++, not a DSL and not a compiler. Documentation structure is the substantive content: a mathematical basis, CK-Tile, and specifically 'CK Tile intrawave and interwave scheduling', which is the AMD analogue of NVIDIA's warp scheduling and is where a CDNA-specific kernel actually differs from a CUDA port. Now developed inside the ROCm/rocm-libraries super-repo; the standalone ROCm/composable_kernel repo is marked '[DEPRECATED] Moved to ROCm/rocm-libraries repo' with develop kept as a read-only mirror, latest release rocm-7.2.4. Cited for category=template-library, template-instantiation strategy, and the intrawave/interwave distinction.