Improve performance a bit by optimizing away

Improve performance a bit by optimizing away
unnecessary system calls and going to a block size of at least
8192 (on normal hosts, anyway).  This improved performance 5% on my
Debian stable host (2.4.27 kernel, x86, copying from root
ext3 file system to itself).
Include "buffer-lcm.h".
(copy_reg): Omit last argument.  All callers changed.
Use xmalloc to allocate rather than trusting alloca
(which is unwise with large block sizes).
Declare locals more locally, if possible.
Use uintptr_t words instead of int words, for a bit more speed
when looking for null blocks on 64-bit hosts.
Optimize away reads of zero bytes on regular files.
In the typical case, insist on 8 KiB buffers, at least.
Avoid unnecessary extra call to fstat when checking for sparse files.
Avoid now-unnecessary cast to off_t, and "0L".
Avoid unnecessary test of *new_dst when checking for same owner
and group.