SSA IL

[Matthew Fluet]

O.k., I can run through all of the regressions with all of the CPS passes
turned off (except for constantPropagation (Array_array0), polyEqual
(MLton_equal)) and compiling -g and only fail on 5 benchmarks: 

smith-normal-form -- need flatten to eliminate the intervening allocation.

(Honestly, I think we should go with the version of IntInf's that does
it's own allocation;  IntInf_arithops are labeled as Functional, so I
don't buy the claim that flattening will always fix the problem; seems
like we could intersperse some allocations in there if we wanted.) 

world5 -- needs a CPS shrink towards the end to turn non-tail calls into
tail calls, otherwise we exhaust stack space

mutex, thread2, prodcons -- don't see to terminate; mutex works o.k. when
not compiled -g; could be they are just very inefficient.