viewer_manifest is run twice by CMakeLists.txt: once with --actions=copy and
once with the default actions copy and package. The fact that we (try to) move
the Mac viewer executable both times has confused things on a number of
occasions. Currently it's bollixing our attempt to run dsymutil on the
executable so we can package up the Mac symbols for both codeticket and
BugSplat.

Introduce DarwinManifest.is_rearranging() that tests whether either "package"
or "unpacked" is in self.args["actions"], echoing an earlier test. Make
several things conditional on that, notably moving the executable.

If this theory is correct, setting VIEWER_APP_EXECUTABLE to VIEWER_BINARY_NAME
and then referencing VIEWER_APP_EXECUTABLE instead of VIEWER_BINARY_NAME
confuses CMake as to the filename involved. <eyeroll/>

Seems VIEWER_SYMBOL_FILE arrives from build.sh as a relative pathname, so
passing that pathname to tar when we run tar with WORKING_DIRECTORY confuses
it. But if we use tar's -C switch, we can achieve the effect we want (no
leading directory prefixes in the tarball) without WORKING_DIRECTORY. Thing
is, we have to run the desired directory through cygpath first, which is why
we went with WORKING_DIRECTORY in the first place.

This is a separate step from generating and posting BugSplat symbols, since
BugSplat needs the executable along with the symbols, and we don't need to
consume that space in a symbols tarball.

Move Mac BugSplat symbol generation logic to CMake land, the same general area
where Breakpad symbols are generated. Add stanzas to pack up the usual tarball
for Windows and Mac.

Remove the build.sh test that suppressed uploading the symbols tarball for
BugSplat builds.

instead of relying on both indra/newview/CMakeLists.txt and build.sh
generating the same file pathname.

Make build.sh set VIEWER_SYMBOL_FILE (instead of symbolfile) in pre_build, and
pass it to autobuild configure via -D switch. Then the uploads stanza can just
use VIEWER_SYMBOL_FILE instead of performing its platform-sensitive case
statement right there.

Introduce VIEWER_SYMBOL_FILE CMake cache variable, default empty string.

Make indra/newview/CMakeLists.txt generate_breakpad_symbols logic conditional
on VIEWER_SYMBOL_FILE being non-empty, as well as everything else. Eliminate
local set(VIEWER_SYMBOL_FILE) directives.

Define the CMake cache variable, with empty string as its default.

Make build.sh pass the BUGSPLAT_DB environment variable as a CMake
command-line variable assignment.

Change CMake 'if (DEFINED ENV{BUGSPLAT_DB})' to plain 'if (BUGSPLAT_DB)'.

Make CMake pass new --bugsplat switch to every one of SIX different
invocations of viewer_manifest.py.

Give llmanifest.main() function an argument to allow supplementing the base
set of command-line switches with additional application-specific switches.

In viewer_manifest.py, define new --bugsplat command-line switch and pass to
llmanifest.main(). Instead of consulting os.environ['BUGSPLAT_DB'], consult
self.args['bugsplat'].

This required reordering certain operations during Mac viewer startup. Split
llappviewermacosx.cpp's initViewer() function into constructViewer() (which
instantiates LLAppViewerMacOSX) and initViewer() (which calls
LLAppViewerMacOSX::init()).

llappdelegate-objc.mm's applicationDidFinishLaunching override now calls
[BugsplatStartupManager start] between constructViewer() and initViewer(): we
want constructViewer() to have set up the logging subsystem so we can log the
actions of BugsplatStartupManagerDelegate override methods, but otherwise we
want BugsplatStartupManager in place as early as possible to catch any early
crashes. Besides, initViewer() ends up overwriting the static_debug_info.log
on which we depend for the *previous* run's crash metadata.

Move the code that initializes the pathname of the static_debug_info.log file
from LLAppViewerMacOSX::init() to the LLAppViewerMacOSX() constructor, since
BugsplatStartupManagerDelegate override methods need to read (the previous
run's) file.

Add code to applicationLogForBugsplatStartupManager override to set new
BugsplatMac 1.0.6 properties userName and userEmail.

Don't log empty fields from static_debug_info.log if we couldn't read it.

Introduce CrashMetadata, an LLSingleton in llappviewermacosx.cpp, declared in
llappviewermacosx-for-objc.h and accessed by the various
BugsplatStartupManagerDelegate override methods. CrashMetadata is populated by
reading the previous (presumably crashed) run's static_debug_info.log file.

This replaces the previous getOldLogFilePathname(), getFatalMessage() and
getAgentFullname() functions. To extend that suite for additional metadata,
not only would we have to keep adding new free functions, but we'd have to
keep rereading the static_debug_info.log file.

Override the new applicationKeyForBugsplatStartupManager,
defaultUserNameForBugsplatStartupManager,
defaultUserEmailForBugsplatStartupManager methods to extract relevant fields
from CrashMetadata. Change applicationLogForBugsplatStartupManager and
attachmentForBugsplatStartupManager to do the same.

Enhance llviewerregion.cpp to update the static_debug_info.log file every
time we enter a new region.

Also use the LLOSInfo information for platform rather than simply Windows32 or
Windows64.

Just put the static_debug_info.log file in the parent logs directory.

Also update that static_debug_info.log file with "FatalMessage" key taken from
LL_ERRS() message string.

I think the intention of (sDumpDir.rbegin() == mDirDelimiter.rbegin()) was to
test whether sDumpDir endsWith(mDirDelimiter). But those iterators will never
be equal. Instead, use LLStringUtil::endsWith().

On Windows, where 'path' might be separated either with '/' or '\', the latter
breaks unless all path separators are in fact the os.sep character '\'. While
it would be possible to code something fancy with os.sep and os.altsep,
testing the latter for None, much simpler to let os.path.split() handle it.