This evidently makes all the difference as to whether the app is considered
launchable.

Specifically, Second Life.app is now mostly just a wrapper. Its Contents/
Resources contains nested Launcher.app (the VMP) and Viewer.app (the viewer
itself). Most of what used to be in the top-level Second Life.app has been
relocated to the embedded Viewer.app. VMP stuff has of course been extracted
to Launcher.app. The top-level Second Life.app executable is now a tiny script
that runs Launcher.app. This structure permits different icons and different
Dock flyover text for the launcher and the viewer, hopefully ameliorating a
certain amount of user confusion about the dual icons.

This requires a corresponding VMP change: on macOS, the VMP must now find both
its resources and the viewer executable by walking up from Launcher.app and
down again into its sibling Viewer.app.

Since Dock flyover text is determined by the embedded app names, allow Product
to change these at will. That means we should be able to tweak exactly one
variable assignment to change either of those embedded app names, without
having to chase down other references scattered throughout the source repo.

For that reason, create top-level trampoline SL_Launcher script dynamically:
it must reference the launcher app by name. That means we must also perform
(the equivalent of) chmod +x on that generated script.

The one mystery surrounding this restructuring is that without a top-level
Frameworks symlink pointing to the embedded Viewer.app's Frameworks directory
(where CEF lives), CEF refuses to start: no splash screen, no MoP. Perhaps we
can fix that someday.

Use Python's bundled plistlib to generate Info.plist files for the embedded
applications.

Reorganize stray code stanzas to try to help the structure of the code more or
less resemble the structure of the desired result.

Add ViewerManifest.relpath() method to determine the relative path from a
specified base to the target path. If base omitted, assumes get_dst_prefix()
-- handy for creating symlinks. Determining exactly the right number of
os.pardir instances to concatenate into the relative pathname for a symlink
(or an install_name_tool stamp) was tedious, fragile and unobvious, difficult
to desk-check. Using relpath() should make all that more robust.

Migrate symlinkf() from free function to ViewerManifest method, refactoring
into _symlinkf_prep_dst() and _symlinkf(), adding relsymlinkf(). This lets us
add convenience features such as prepending get_dst_prefix() to the dest (the
place where we want to create the symlink), defaulting dest to the basename of
target and ensuring that the parent of that dest already exists -- as with
LLManifest.path(). Moreover, since it makes no sense whatsoever to create an
absolute symlink to some path on the build machine, relsymlinkf() creates
every symlink relative to dirname(dest). That, in turn, lets us eliminate a
certain amount of boilerplate around existing calls. (Also, since we now
ensure the parent directory exists, scrap the logic to diagnose "nonexistent
parent directory.")

Make llmanifest.LLManifest.run_command() not pass shell=True to subprocess,
thereby permitting (requiring) the list form rather than the string form.
Change all existing calls to list form. This makes calls more readable, for
two reasons. First, many of the arguments are taken from script variables;
these can simply be dropped into the list instead of indirecting through
string interpolation. Second, it eliminates the need to manually escape
individual arguments, since subprocess promises to honor the distinction
between list elements.

Also fix LLManifest.put_in_file() to ensure the containing directory exists.

Consolidate some viewer_manifest.py redundancy, e.g. copying the same set of
ten DLLs from either of two directories depending on Release vs. Debug.

The LLManifest.copy_action() method ensures that the destination directory for
any file copy exists before trying to copy the file, specifically so we don't
have to clutter the business logic with tests and explicit directory creation.
Remove redundant os.makedirs() stanzas.

Introduce helper classes to manage paired initProfile() / finishProfile()
calls and gBenchmarkProgram.bind() / unbind() calls.

Make TextureHolder a class instead of a struct.

Per Henri Beauchamp, since gpu_benchmark() takes a very early exit if
(!gGLManager.mHasTimerQuery), subsequent tests of mHasTimerQuery are
redundant. Remove.

One of those tests controls the busted_finish bool, which can never become
true. Remove that and all tests on it.

A classic-C array doesn't destroy its individual elements, but a std::vector
does. Use a std::vector<LLRenderTarget> for dest, so each LLRenderTarget will
be destroyed. ~LLRenderTarget() calls its release() method.

Ruslan tracked the observed crash to assignments (to create a dummy triangle)
through an LLStrider<LLVector3> obtained from getVertexStrider(). When
getVertexStrider() returns false, produce a warning and just skip the rest of
the benchmark test.

The one bit of explicit cleanup apparently required by that early exit is a
call to LLImageGL::deleteTextures() to match the preceding generateTextures()
call. Wrap both in a new TextureHolder class whose destructor takes care of
cleanup. The only other references to the corresponding U32 array are a couple
calls to LLTexUnit::bindManual(); add a bind() method to support that.

Also fix apparent bug in the LL_DARWIN special case for "improbably high and
likely incorrect": the code assigned -1.f (the "couldn't compute" value) to
gbps, overlooking the fact that gbps is unconditionally recomputed below. In
the "likely incorrect" stanza, simply return -1.f instead.

In case of buff->getVertexStrider(v) return false it mean that glMapBufferRange() return NULL
The next three lines can be the reason of this crash.