Major build cleanup.

[adam-azarchs]

Fix the Makefile dependencies so things get rebuilt when they should be.

Only produce one .so file, which knows how to run both 2d and 3d.

Break out main and the I/O methods into a separate cpp file, which is
not used for the python extension module.

Do not create pointless objects. The methods are all static. TSNE is
now a namespace, not a class, and the exported run method is wrapped
in extern "C" to make linkage easier.

Minor cleanup of includes / using statements.

Because I just couldn't help myself, remove a couple places where
new/delete were being used for no good reason.

Get rid of a few places where NDIMS (the build time constant) and a
run-time variable, which are supposed to have the same value, were both
being used in a method. This means flowing the template a bit further
up the tree. There's lots more places this could be done - this only
addresses the cases where not doing so could lead to errors if the
template variable was inconsistent with the run-time variable.

Kill off the last pointless vestiges of the cblas dependency.

Use hidden visibility where appropriate - only export the run() method.
This cuts the size of the .so and allows the compiler to inline more
aggressively, especially since we've also turned on LTO.

Together, all of these changes result in a build with a single 901k
extension module, instead of two 1.9MB extension modules.

[nlhepler]

I am happy with where this is, I think the only thing I could pessimistically think of is a test (with fixed seed) that equates the result with a known quantity.

[adam-azarchs]

I moved the consistency test to the start of the commit chain and remade it after disabling -ffast-math from the build. The new HEAD commit still passes. There were some differences introduced by different treatment of duplicate points which have been fixed now.

Then I turned -ffast-math back on since it seems to still pass the tests.

[adam-azarchs]

On a somewhat larger dataset, which took around a minute on marsoc, running on hydra we see

key              change    before       after
     walltime   -21.1 %        35          28
 max_walltime   -21.1 %        35          28
        utime   -21.5 %        34          27
    max_utime   -21.5 %        34          27
        stime    -5.5 %         1           1
    max_stime    -5.5 %         1           1
          rss   -------    628028      627996
      max_rss   -------    623012      622860
     max_vmem   -------   1294012     1291956

which is roughly consistent with the improvement I expected to see.