Otherwise, systems that don't support it will treat it as a name for a
global `struct ntp_msg` variable.

This partially reverts commit a7577667d6.
The musl C library does not use an internal cdefs.h header. As a result,
the definitions in sys/cdefs.h have no effect. This breads the build as
follows:
In file included from adjfreq_linux.c:23:0:
../src/ntpd.h:448:8: error: expected ‘=’, ‘,’, ‘;’, ‘asm’ or ‘__attribute__’ before ‘void’
__dead void fatal(const char *, ...)
^~~~
../src/ntpd.h:450:8: error: expected ‘=’, ‘,’, ‘;’, ‘asm’ or ‘__attribute__’ before ‘void’
__dead void fatalx(const char *, ...)
^~~~
Refer the related musl FAQ about cdefs.h:
http://wiki.musl-libc.org/wiki/FAQ#Q:_I.27m_trying_to_compile_something_against_musl_and_I_get_error_messages_about_sys.2Fcdefs.h
Keep the compiler attribute definition.
Remove also from noinst_HEADERS.

Solaris and illumos based distributions) which don't have it.

This also removes the need for __weak_alias shims

I could not find a lot of precedence for this, because most time
daemons do not actually look at the value of olddelta. Account for
olddelta getting stuck at 1ms, and for a NULL value of delta being
treated as an error condition.

Due to some circular dependencies on Solaris, this needs to be defined
as a macro. We should just remove SA_LEN usage here next.

If sa_family is not set, sa_len should be zero.

I removed all of the code in fake-sensors, I don't think Henning wants
all the credit for an empty file.

TODO: convert u_int_* types to uint_*