Signed-off-by: Max Filippov <jcmvbkbc@gmail.com>

Instead, write the @soname bit separately.  I'm not sure why this was done
like that in the first place.  I may have cut'n'pasted this from RHEL
patches for dlopen support, where these sorts of hacks were necessary.

This needs to be checked in three steps, adding dependent libraries one
after another.  It also isn't necessary to have three variables to keep
track of what's one conceptual unit, so drop libunwind_{ptrace,arch}_LIBS.

Include a test suite case.
Describe the extension in ltrace.conf.5.
Mention it in NEWS.

There's no similar agreed-upon convention for oct, and displaying which
bits are set in a floating number like what bitvect does is certainly not
helpful.  Hence, remove the corresponding TODO item.  Add a different
(unrelated) one instead.

- add symbolic names for signals and use bit vectors to show sigset_t.
- add GNU extensions sig{and,or,isempty}set
- sort the list of updated prototypes in NEWS

The reason for this is that there's no agreed-upon term for "hexadecimal"
that wouldn't be super awkward to type.  So instead keep "hex", but
introduce "oct" for consistency.

Add a brief note to ltrace.1 as well.

This option doesn't exist anymore.

- "void main(void);" accepted without warning to mean "void main();"
- "itn" accepted without warning to mean "int"
- "void main(void, xxx);" accepted with warning to mean
  "void main(hide(int), xxx);".  This gets warning, because shipped
  ltrace.conf never contained this idiom.

Update NEWS with explanation of the above.

Add more detail to some points.
Shuffle entries around.

This should be used for tracking user-visible changes.  I retroactively
filled in entries from 0.6 on (inclusive).