Replace our own handcrafted I/O clustering with the generic write_cache_pages helper. While the old code would add any additional page in an existing mapping, the new code iterates over the pages, either adding them to a previous mapping if it fits or otherwise gets a new one. This has survived xfsqa for small and 4k blocksize on x86.