Rewrite --init-storage to dump a compiled-in blob
instead of firing off the entire save snapshot machinery.

This should simplify bootstrap in presence of system
spaces:
there's going to be no need to implement the
code which creates system, they will be created
simply by recovery from a snapshot.

Two significant changes (both arguable, since this patch is about
taste by a large measure):

1) Split tuple_seek() into two: tuple_rewind()
and tuple_seek().

tuple_rewind() simply initializes the iterator and
sets it to 'before-first' position.
tuple_seek() returns a field, and sets the iterator
to the position after the returned field.

Previosly there was only tuple_seek(), which required
tuple_next() to get the field, and it was confusing whether
or not tuple_next() after tuple_seek() is required.
Indeed, the very code I chagned sometimes would
call tuple_next() after tuple_seek() and sometimes would not.

2) Move the return value of the iterator outside
of iteration state.

The advantage of "binding" the out parameters of the iterator to
iterator state is that you bind once, and don't need to supply out
parameters into every call to tuple_next() function. It also makes
it easy to add extra out parameters for bind in the future (field
data type, flags, etc).

This makes "binding" technique popular in database driver
APIs. Binding makes API more stable, which is also
important in drivers.

In this case, however, binding adds for extra lines of code
and (possibly) few extra instructions. Plus, since
this is an internal API which we can change any day,
its stability is not as important. So it seems that for
now we're better of with an API which is a bit
more concise/efficient, and can switch to using binding later,
if needed.

Don't load corrupted tuples.

g++ (not gcc) works correctly when dealing with
extern __inline and __extern_inline constructions, thus
this hack is not necessary now.

The patch which removes the CPack-generated spec
and adds a proper spec for client, server and library
was done by Dmitry Simonenko.

Hand-port the patch to the master, since it's gone too
far away from the origin and merging it ends up in a mess.

Code review of cjson branch by Dmitry E. Oboukhov.
Hand-picking the changes since the original branch
is based on import of lua-cjson into Tarantool
revision tree, and this branch is based on a git
submodule.

This patch implements
https://blueprints.launchpad.net/tarantool/+spec/lua-cjson

This messes the old "core/module" paradigm quite a bit, but Box,
while a neat name, was confusing people.

tarantool_box is preserved as the daemon name.

libreadline already goes with libhistory compiled in,
no reason to add it to the dynamically linked library
list.
Apparently, explicitly listing libhistory leads to
the dl loader pre-loading it instead of the compiled-in
version, which, in turn, leads to a crash on the first
history search in the command line client.

The fix is to remove cmake dependency on libhistory
as well as on libncurses, which we don't need either
(always require libtermcap).

twistd by default starts with umask 077.
This is why buildbot has umask setting.

rpm by default takes file and directory umask
from the build directory.

When building tarantool rpm under buildbot,
twistd default umask was inherited. Thus
all paths were created in rwx-- mode.

Now add to this cmake bug CMake bug 0012864.
That is, by default, cmake cpack generator
adds %dir directive for every intermediate
dir into the spec.

E.g. if there is a destination file /etc/rc.d/foo,
the spec gets
%dir /etc
%dir /etc/rc.d

Together with the default dir access mask of
twistd this nicely messed up our rpm install: an
install of Tarantool 1.4.8-59 rpm for CentOS would change
/etc and /usr access mode to 700.

Two things are done about that: Tarantool no longer
uses a cpack generated cpeck file. Extraneous %dir
directives are removed from the spec file. This
should minimize the damage of a broken twistd config,
should it happen again.

All buildslaves are reconfigured and restarted with umask 022.

When make test is run with gcov, coverage output files are
created. On the second run, when they are already there,
the server fails to start.