Revert r149083 which is not the direction we're going in the Clang

driver based on discussions with Doug Gregor. There are several issues:
1) The patch was not reviewed prior to commit and there were review comments.
2) The design of the functionality (triple-prefixed tool invocation)
   isn't the design we want for Clang going forward: it focuses on the
   "user triple" rather than on the "toolchain triple", and forces that
   bit of state into the API of every single toolchain instead of
   handling it automatically in the common base classes.
3) The tests provided are not stable. They fail on a few Linux variants
   (Gentoo among them) and on mingw32 and some other environments.

I *am* interested in the Clang driver being able to invoke
triple-prefixed tools, but we need to design that feature the right way.
This patch just extends the previous hack without fixing the underlying
problems with it. I'm working on a new design for this that I will mail
for review by tomorrow.

I am aware that this removes functionality that NetBSD relies on, but
this is ToT, not a release. This functionality hasn't been properly
designed, implemented, and tested yet. We can't "regress" until we get
something that really works, both with the immediate use cases and with
long term maintenance of the Clang driver.

For reference, the original commit log:
Keep track of the original target the user specified before
normalization. This used to be captured in DefaultTargetTriple and is
used for the (optional) $triple-$tool lookup for cross-compilation.
Do this properly by making it an attribute of the toolchain and use it
in combination with the computed triple as index for the toolchain
lookup.

llvm-svn: 149337