Add test that exercises all bfd architecture, osabi, endian, etc. combinations

Add test that exercises all bfd architecture, osabi, endian, etc. combinations

This adds a test that exposes several problems fixed by earlier
patches:

#1 - Buffer overrun when host/target formats match, but sizes don't.
     https://sourceware.org/ml/gdb-patches/2016-03/msg00125.html

#2 - Missing handling for FR-V FR300.
     https://sourceware.org/ml/gdb-patches/2016-03/msg00117.html

#3 - BFD architectures with spaces in their names (v850).
     https://sourceware.org/ml/binutils/2016-03/msg00108.html

#4 - The OS ABI names with spaces issue.
     https://sourceware.org/ml/gdb-patches/2016-03/msg00116.html

#5 - Bogus HP/PA long double format.
     https://sourceware.org/ml/gdb-patches/2016-03/msg00122.html

#6 - Cris big endian internal error.
     https://sourceware.org/ml/gdb-patches/2016-03/msg00126.html

#7 - Several PowerPC bfd archs/machines not handled by gdb.
     https://sourceware.org/bugzilla/show_bug.cgi?id=19797

And hopefully helps catch others in the future.

This started out as a test that simply did,

 gdb -ex "print 1.0L"

to exercise #1 above.

Then to cover both 32-bit target / 64-bit host and the converse, I
thought of having the testcase print the floats twice, once with the
architecture set to "i386" and then to "i386:x86-64".  This way it
wouldn't matter whether gdb was built as 32-bit or a 64-bit program.

Then I thought that other archs might have similar host/target
floatformat conversion issues as well.  Instead of hardcoding some
architectures in the test file, I thought we could just iterate over
all bfd architectures and OS ABIs supported by the gdb build being
tested.  This is what then exposed all the other problems listed
above...

With an --enable-targets=all, this exercises over 14 thousand
combinations.  If left in a single test file, it all consistenly runs
in under a minute on my machine (An Intel i7-4810MQ @ 2.8 MHZ running
Fedora 23).  Split in 8 chunks, as in this commit, it runs in around
25 seconds, with make -j8.

To avoid flooding the gdb.sum file, it avoids calling "pass" on each
tested combination/iteration.  I'm explicitly not implementing that by
passing an empty message to gdb_test / gdb_test_multiple, because I
still want a FAIL to be logged in gdb.sum.  So instead this puts the
internal passes in the gdb.log file, only, prefixed "IPASS:", for
internal pass.  TBC, if some iteration fails, it'll still show up as
FAIL in gdb.sum.  If this is an approach that takes on, I can see us
extending the common bits to support it for all testcases.

gdb/testsuite/ChangeLog:
2016-12-09  Pedro Alves  <palves@redhat.com>

	* gdb.base/all-architectures-0.exp: New file.
	* gdb.base/all-architectures-1.exp: New file.
	* gdb.base/all-architectures-2.exp: New file.
	* gdb.base/all-architectures-3.exp: New file.
	* gdb.base/all-architectures-4.exp: New file.
	* gdb.base/all-architectures-5.exp: New file.
	* gdb.base/all-architectures-6.exp: New file.
	* gdb.base/all-architectures-7.exp: New file.
	* gdb.base/all-architectures.exp.in: New file.