This adds -lblas to the lapack library flag, because with the last
update the toolchain on Belize is more strict about implicit library
inclusions.

Also, the library flags have been re-ordered in the general makefile,
going (roughly) from most specific to least.

In the system-specific build setting scripts, the OPTIM_LDFLAGS
variable was generally empty.  As it turns out, at least for
Itanium systems, this disabled a key optimization or two, leading
to the emergence of denormalized floats (which, on these systems,
are handled in software).  Testing shows that including -O3 at
the link step should keep this optimization enabled (really just
a modification of the default floating-point environment).  Since
this change is unlikely to break anything, the fix for all systems
has been to copy the $OPTIM_CFLAGS variable to $OPTIM_LDFLAGS.

Using code lifted from internal_wave_dl, the benchmark case now reads
/proc/self/statm on a per-process basis in order to get the resident
set size.  This solution is probably -not portable- outside of Linux
systems; when the file is not present the execution will likely die with
an assertion failure.

Additionally, no effort is made to properly account for shared memory --
a hypothetical run that uses 1MB/proc of local memory and 99MB/proc of
shared memory will report a total of 100MB * procs memory usage, rather
than 1 * procs + 99MB.  A general solution here is difficult because it
relies on accounting for shared memory once-per-system (and nhost > 1
for cluster configurations).

The rectangular-matrix solve inside the GMRES code is computed with the
LAPACK function DGELSD.  This function has a 'workspace query' in order
to give the proper allocation of the workspace array.  As it turns out,
the -integer- work array (as opposed to the floating-point one) is of a
fixed size, and it can be allocated ahead of time.  Not doing so caused
problems with MKL on SciNet, since even for the query the intel version
of LAPACK wrote to this array (causing a segmentation fault).

In a Valgrind check, several usages of uninitialized memory in the
multigrid code were being caught.  This was related to the check for
boundary-condition indefiniteness/validity at the coarsest level.  In
effect these errors were harmless, but fixing them will make future
Valgrind checks cleaner.

Also turned mapped_grid from switch to full boolean, so that it makes
more sense in a configuration file (that doesn't support "switch"-type
parameters as such, turning them into boolean).

Remaining changes:
  *) Properly account for non-zero initial time, by setting the first
plot sequence number appropriately
  *) Add a generate_automatic_grid method to write out [xyz]grid files
from the constructor, without overloading init_tracer
  *) Add 3d_perturbation parameter for restart-in-three-dimensions cases
  *) Test on a few of the already-existing cases

Also included an add_option speicalization for the string parameter
type.  It's convenient to allow for const char * as a default value for
these parameters, but that poses a problem for the templated add_option
function because of the lack of automatic conversion to const string.
By including a specialized overload of the template (with
implementation in Options.cpp), this problem is fixed.

This library uses boost::program_options, which provides a command-line
parser that also reads key=value - style configuration files.
Options.c/hpp provides a simplified interface for cases-code, which
should make option-definition more natural.

On systems using the MKL library, including the MKL (via $(LAPACK_LIB))
brings in the MKL fftw-wrapper bindings, rather than fftw itself.  This
is problematic on account of the fact that the MKL fftw wrapper doesn't
implement the parts of the library that SPINS actually uses; this leads
to an assertion failure on startup.

Also tweaked make_deps to properly install AMD and to
pick up on SYSTEM_CFLAGS (necessary for 32/64-bit
switch on plata).  Consequently, added SYSTEM_CXXFLAGS
since -Wno-deprecated is inappropriate for pure-C