xlog: rework writer API

The goal of this patch is to make the xlog writer API less confusing.
It does the following changes:

 - Rename xlog_rename to xlog_materialize. The function strips the
   .inprogress suffix from an xlog file name. Usually it's called after
   the xlog was sealed (by writing the EOF marker) and closed but it may
   also be called before xlog_close in case the xlog file is going to be
   appended to (used in WAL).

 - Make xlog_sync private. It's called by xlog_close so there's no point
   in calling it manually. Make xlog_close fail if it fails to sync the
   file.

 - Make xlog_close flush the buffer so that the caller doesn't need to
   call xlog_flush manually before closing an xlog file.

 - Drop the reuse_fd argument of xlog_close. Instead, introduce a new
   function xlog_close_reuse_fd that works exactly like xlog_close
   except it doesn't close the fd and returns it in the out argument.

 - Introduce xlog_discard. It's supposed to be called on an incomplete
   xlog file to close its fd and unlink it.

Rework all the places where we write xlog files in accordance with the
new API. Note about memtx snapshot writer: use xlog_materialize and
xlog_discard instead of coio_rename and coio_unlink.

The new API (especially the xlog_discard helper) facilitates
introduction of thorough (secure) file deletion because it significantly
reduces the number of places in the code where xlog files are deleted.

Needed for tarantool/tarantool-ee#540

NO_DOC=refactoring
NO_TEST=refactoring
NO_CHANGELOG=refactoring