Rebuild windows apr to fix CRT dll hell.  Also upgrading to apr-1.3.8/apr-util-1.3.9 in the process for DEV-38196 vulnerability.

Added CRT assembly check to viewer_manifest.py.
twiddled test_win32_manifest.py for ease of use.

In the viewer-2 code base, the "tos" case is detected inline, a sibling of the
other types of login failure, and doesn't reach the notification. Our new
logic needs to detect "tos" as well.
Also, the case of process_login_success_response() returning false was
inconsistent, attempting a notification with an empty string.

LLFloater's destructor calls LLFloaterReg::removeInstance() with its own name
and key. But for the new LLFloaterTOS invocation, we pass a key that's an LLSD
map. removeInstance() critically depends on LLFloater::KeyCompare::equate() --
but equate() never considered a non-scalar LLSD key value. Fortunately
llsdutil.h already provides a deep-equality function for LLSD: llsd_equals().
Making equate() trivially call llsd_equals() fixes the crash on TOS cancel.

The viewer-2 onCancel() pops up a "MustAgreeToLogIn" notification. Make ours
do the same.

Neither lllogininstance.cpp nor lllogininstance_test.cpp need llfloatertos.h
any more, since LLLoginInstance talks to LLFloaterTOS only via LLFloaterReg
and LLEventPumps. However, both sources depended on llfloatertos.h dragging in
llnotifications.h, so include that explicitly instead of llfloatertos.h.

Because the details of RunBuildTest.cmake versus run_build_test.py had to be
changed in so many different places, introduce LL_TEST_COMMAND CMake macro (in
LLTestCommand.cmake) to encapsulate construction of the actual command line.
Use LL_TEST_COMMAND in LL_ADD_PROJECT_UNIT_TESTS, LL_ADD_INTEGRATION_TEST, the
big indra/test monolith and the various LslCompilerMacros.
Fix run_build_test.py to pass through the test executable's own options (e.g.
--touch, --output) without inspection. Defend it against the case when the
platform-specific library path environment variable doesn't yet exist. Make it
report errors only on nonzero test-program rc.
Remove RunBuildTest.cmake.

qtwebkit4.dll etc aren't being staged to the sharedlibs dir like everything else.

RunBuildTest.cmake can't handle pathnames containing spaces.
run_build_test.py accepts an arbitrary number of individually-quoted
command-line arguments, passing each through to Python's subprocess.call().

the "Copying staged dlls" pre-build step for newview got moved to be a pre-build step for
create_app_config_file and create_app_config_file now depends on stage_third_party_libs.

reviewed by Nat.

The problem arose because we were setting LL_COMMON_BUILD in
llcommon/CMakeLists.txt, not only for the library build itself but also for
its LL_ADD_INTEGRATION_TEST tests. This told all the headers compiled into the
INTEGRATION_TEST_lllazy executable that the executable was providing all the
llcommon symbols, rather than importing them.
The solution is to switch to the llcommon_EXPORTS symbol automagically defined
by CMake when building the llcommon shared library itself.

Change LLDir_Mac::getLLPluginLauncher() to look in the viewer's executable dir
instead of in the plugins dir. Change viewer_manifest.py's
DarwinManifest.construct() to put SLPlugin in the new location.
SLPlugin is being linked with our new libllcommon.dylib, which self-identifies
as being findable via @executable_path/../Resources/libllcommon.dylib. This
doesn't work from the Resources/llplugin subdir -- the above relative path
ends up looking in the nonexistent Resources/Resources subdirectory. Putting
SLPlugin in the Contents/MacOS directory with the viewer executable solves the
problem.

when you pass in a command string with command-line arguments,
RunBuildTest.cmake attempts to search for a program whose filename is the
entire command line. Uncommented separate_arguments().
Added SHARED_LIB_STAGING_DIR to LL_ADD_INTEGRATION_TEST LD_LIBRARY_PATH.