Fix generation of x86-64 gdbarch with osabi none (PR 22979)

When a 64-bits (x86-64) gdbarch is created, it is first born as a
32-bits gdbarch in i386_gdbarch_init.  The call gdbarch_init_osabi will
call the handler register for the selected (arch, osabi) pair, such as
amd64_linux_init_abi.  The various amd64 handlers call amd64_init_abi,
which turns the gdbarch into a 64-bits one.

When selecting the i386:x86-64 architecture with no osabi, no such
handler is ever called, so the gdbarch stays (wrongfully) a 32-bits one.

My first idea was to manually call amd64_init_abi & al in
i386_gdbarch_init when the osabi is GDB_OSABI_NONE.  However, this
doesn't work in a build of GDB where i386 is included as a target but
not amd64.  My next option (implemented in this patch), is to allow
registering handlers for GDB_OSABI_NONE.  I added two such handlers in
amd64-tdep.c, so now it works the same as for the "normal" osabis.  It
required re-ordering things in gdbarch_init_osabi to allow running
handlers for GDB_OSABI_NONE.

Without this patch applied (but with the previous one*) :

  (gdb) set osabi none
  (gdb) set architecture i386:x86-64
  The target architecture is assumed to be i386:x86-64
  (gdb) p sizeof(void*)
  $1 = 4

and now:

  (gdb) set osabi none
  (gdb) set architecture i386:x86-64
  The target architecture is assumed to be i386:x86-64
  (gdb) p sizeof(void*)
  $1 = 8

* Before the previous patch, which fixed "set osabi none", this bug was
  hidden because we didn't actually try to generate a gdbarch for no
  osabi, it would always fall back on Linux.  Generating the gdbarch for
  amd64/linux did work.

gdb/ChangeLog:

	PR gdb/22979
	* amd64-tdep.c (amd64_none_init_abi): New function.
	(amd64_x32_none_init_abi): New function.
	(_initialize_amd64_tdep): Register handlers for x86-64 and
	x64_32 with GDB_OSABI_NONE.
	* osabi.c (gdbarch_init_osabi): Allow running handlers for the
	GDB_OSABI_NONE osabi.

gdb/testsuite/ChangeLog:

	PR gdb/22979
	* gdb.arch/amd64-osabi.exp: New file.