bfd: don't silently wrap or truncate PE image section RVAs

In PE images section addresses get expressed as addresses relative to
the image base. Therefore the VA of a section must be no less than the
image base, and after subtraction of the image base the resulting value
should fit in 32 bits. (The issue is particularly obvious to notice when
sections, perhaps because of ELF assumptions, get placed at VA 0 by
default. Debugging info sections as well as .comment, when input files
are ELF, are a good example. All such sections need proper mentioning in
the linker script to avoid this warning.)

There are a number of test cases which previously produced bogus images,
yet still declared the test a success. Like done for other tests
already, force a zero image base for these. This then also allows (and
requires) dropping again xfail-s which 39a7b38fac0e ("Fix linker tests
to work with 16-bit targets") had added to ld-scripts/default-script*.d
(originally as skip-s). This also depends on similar adjustments to
testsuite/ld-scripts/map-address.* made by an earlier patch.

For ld-scripts/print-memory-usage.* I suppose xcoff could be dropped
from the exclusion list by suppressing garbage collection, just like
already done in e.g. (as seen in the diff here) ld-scripts/data.*, but I
didn't want to make unrelated adjustments.

ld/testsuite/ld-scripts/default-script3.d

@@ -1,8 +1,6 @@
 #source: default-script.s
 #ld: -defsym _START=0x800 -dT default-script.t
 #nm: -n
-#xfail: {[is_pecoff_format x86_64-*]}
-# Skipped on Mingw64 and Cygwin because the image base defaults to 0x100000000
 
 #...
 0*800 . _START