Recognize DragonFly BSD, and fix what keeps the tests from running on the BSDs

[zakinko]

Abseil does not build on DragonFly BSD, and on the other BSDs its test
suite fails in places that have nothing to do with the code under test.
This fixes those, and reports what is left.

• DragonFly is a FreeBSD descendant but is spelled DragonFly, so
  it missed every FreeBSD branch: GetTID, the auxv type names in
  vdso_support, two config macros, and cctz's _XOPEN_SOURCE, which
  hides the C99 wide functions from libstdc++'s there.
• raw_logging.cc did not list NetBSD and DragonFly among the systems
  with write(), so a FATAL there aborted without its message and every
  death test failed with "died but not with expected error".
• stripping_test could not open its own executable on the BSDs. It
  now uses KERN_PROC_PATHNAME on FreeBSD, DragonFly and NetBSD. OpenBSD
  has no such call, and the tests are skipped there.
• convert_test skipped NaN on Apple by name because "%+f" prints "nan"
  with no sign there; the BSDs do the same, so that is now a probe.
  The hex-float rounding probe gains the case where rounding carries
  into the exponent, which FreeBSD prints as "0x1p+1" for 1.5.
• OpenBSD links with --execute-only, which refuses the data that
  symbolize_test places in .text on purpose; that one test links with
  --no-execute-only there.
• GrowExtremelyLargeTable reserves close to 8 GB of address space and
  now skips itself where RLIMIT_DATA is below that, which is OpenBSD's
  default.

Measured with CMake (ABSL_BUILD_TESTING, googletest at head, Release,
C++17):

NetBSD 11.0     gcc 12.5    241 of 242
FreeBSD 15.1    clang 19    240 of 242
OpenBSD 7.9     clang 19    240 of 242 (a third, clock_interface_test,
                            failed once under load and passes alone)
DragonFly 6.4   gcc 14      240 of 242 (gcc 14 from dports; Abseil
                            requires gcc 10 or later, base has 8.3)

The tests still failing are str_format_convert_test and charconv_test,
and in every case the failure is in the platform's libc, which those
tests use as the reference: FreeBSD, NetBSD and DragonFly do not type
the argument of a positional %F ("%1$*2$F"), and on FreeBSD the
vsnprintf return value then makes the test's helper allocate hundreds
of megabytes at a time, so I ran that test alone under a 4 GB address
space limit; OpenBSD prints nothing for "%#.0o" of zero; DragonFly's
"%lc" encodes a negative wchar_t instead of failing; and nan(3) and
strtod disagree on NaN payloads on FreeBSD, OpenBSD and DragonFly. Each
was confirmed with a small C program, and StrFormat and from_chars give
the C answer in every case. I have left those comparisons in place
rather than skip them; the bugs are for the BSDs to fix, and I will
report them there.

[brad0]

OpenBSD has no such call, and the tests are skipped there.

That is no longer true. OpenBSD now has getexecpath().