lapackers

[Jason Riedy <ejr@cs.berkeley.edu>]

Just a quick heads-up before I forget.  Anyone using a LAPACK
library built via ATLAS will not have:
  1) any of our new routines, or
  2) any of the NaN checking in LU or Cholesky.

The consequence of #1 should immediately stop people from
using said library if they reference QR or use a shared
library version.  The QR factorization routines will be
missing the new Householder generation routines.  These bugs
should manifest as missing symbols related to xLARFP, missing
symbols from the improved QR iteration, or possibly others.

Consequences of #2 will be subtle.  We don't have any NaN
tests, and I doubt if the contributors (MathWorks) use the
ATLAS-LAPACK procedure.

ATLAS can rip apart and rebuild the LAPACK library, replacing
xGETRF, xGETRI, xGETRS, xPOTRF, xPOTRS, xTRTRI, and a few
other routines.  The build routine works off an old list of
object files and not the object files actually present.

wheee....

Jason, who really wants a 70-80% of peak BLAS with a build
procedure that doesn't suck...