Upgrade XGBoost to use BB on all platforms

[staticfloat]

This upgrades the package (and a few stylistic points within the package itself) to use prebuilt binaries instead of requiring the user to compile on their own machine.

[staticfloat]

I've switched this to WIP, because in my testing, Windows doesn't work yet, and I also suspect that there are std::string values that cross the API boundary, which means we need to test out the new BB multiversioning capability to match against std::string ABI version. (EDIT: This was incorrect; this package only uses std::basic_string, so there are no incompatibilities) Hold tight. :)

[staticfloat]

So the problem with Windows is that the binaries immediately crash Julia upon dlopen(). The backtrace is the following:

julia> using Libdl; Libdl.dlopen("xgboost.dll")
warning: Critical error detected c0000374

Thread 1 received signal SIGTRAP, Trace/breakpoint trap.
0x000007fba7a7a598 in ntdll!RtlpNtCreateKey () from /cygdrive/c/Windows/SYSTEM32/ntdll.dll
(gdb) bt
#0  0x000007fba7a7a598 in ntdll!RtlpNtCreateKey () from /cygdrive/c/Windows/SYSTEM32/ntdll.dll
#1  0x000007fba7a7f3f0 in ntdll!RtlpNtCreateKey () from /cygdrive/c/Windows/SYSTEM32/ntdll.dll
#2  0x000007fba7a25f05 in ntdll!RtlSetCurrentEnvironment () from /cygdrive/c/Windows/SYSTEM32/ntdll.dll
#3  0x000007fba54c11bc in msvcrt!free () from /cygdrive/c/Windows/system32/msvcrt.dll
#4  0x000000006fcb366d in libstdc++-6!_ZNSs6assignERKSs () from /home/Administrator/dist/bin/libstdc++-6.dll
#5  0x0000000029d96f9d in dmlc::data::LibSVMParserParam::__MANAGER__() () from /cygdrive/c/Users/Administrator/.julia/packages/XGBoost/VnexW/deps/usr/bin/xgboost.dll
#6  0x0000000029f0559e in _GLOBAL__sub_I_data.cc () from /cygdrive/c/Users/Administrator/.julia/packages/XGBoost/VnexW/deps/usr/bin/xgboost.dll
#7  0x0000000029da4ed5 in __do_global_ctors () at /workspace/srcdir/mingw-w64-v6.0.0/mingw-w64-crt/crt/gccmain.c:67
#8  0x0000000029d2129e in __DllMainCRTStartup (hDllHandle=0x29d20000, dwReason=1, lpreserved=0x0) at /workspace/srcdir/mingw-w64-v6.0.0/mingw-w64-crt/crt/crtdll.c:193
#9  0x000007fba79ab7df in ntdll!RtlAllocateAndInitializeSid () from /cygdrive/c/Windows/SYSTEM32/ntdll.dll
#10 0x000007fba79bd288 in ntdll!RtlLocaleNameToLcid () from /cygdrive/c/Windows/SYSTEM32/ntdll.dll
#11 0x000007fba79bc817 in ntdll!RtlLocaleNameToLcid () from /cygdrive/c/Windows/SYSTEM32/ntdll.dll
#12 0x000007fba79be5a6 in ntdll!RtlLocaleNameToLcid () from /cygdrive/c/Windows/SYSTEM32/ntdll.dll
#13 0x000007fba79be966 in ntdll!RtlLocaleNameToLcid () from /cygdrive/c/Windows/SYSTEM32/ntdll.dll
#14 0x000007fba79d1c47 in ntdll!LdrLoadDll () from /cygdrive/c/Windows/SYSTEM32/ntdll.dll
#15 0x000007fba4bc7c4d in LoadLibraryExW () from /cygdrive/c/Windows/system32/KERNELBASE.dll
#16 0x000000006b5af0ba in jl_dlopen (filename=filename@entry=0x29d1eaf0 "xgboost.dll", flags=flags@entry=68)
    at /home/Administrator/buildbot/worker/package_win64/build/src/dlload.c:85
#17 0x000000006b5af2b0 in jl_load_dynamic_library (modname=0xd675d58 "xgboost.dll", flags=68, throw_err=701623824)
    at /home/Administrator/buildbot/worker/package_win64/build/src/dlload.c:189
#18 0x00000000030e7a19 in julia_#dlopen#3_14362 () at C:\cygwin\home\Administrator\buildbot\worker\package_win64\build\usr\share\julia\stdlib\v1.1\Libdl\src\Libdl.jl:109
#19 0x0000000028d20e0a in ?? ()
Backtrace stopped: previous frame inner to this frame (corrupt stack?)

[staticfloat]

Alright, I'm not certain why I thought std::basic_string values were somehow exempt from the cxx11 string ABI problems, but they aren't. Building string-abi-specific tarballs fixes this, and this package passes tests on linux64, mac64 and win64, so I am taking off the WIP tag. :)

[aviks]

Thanks Elliot, this is exciting. Do you know if these binaries are built with openmp?

[staticfloat]

All except for MacOS are. This is because on MacOS, we are using clang to build Julia, and getting OpenMP to work with clang is apparently complicated, I haven't looked into it too much. Using gcc on MacOS will cause us to mix C++ stdlibs, and I don't think that's a good thing, so we're sticking with non-OpenMP on macOS. Is that reasonable?

[aviks]

Is that reasonable?

eminently.