The test case loop.erl shows that there is a problem with certain
singleton key types. Here the internal representation toggles between
 #{a | b => ...} and #{a => ..., b => ...}

The choice is to turn #{a | b => ...} into #{a => ..., b => ...} early
(t_from_form()). The aim is to keep as much info as possible (in
pairs). However, including complex singleton keys (tuples, maps) in
this scheme is potentially too costly, and a bit complicated. So one
more choice is made: let atoms and number (and nothing else) be
singleton types, and let complex keys go into the default key.

HiPE's range analysis would not update the arguments of a callee when
the result of the call was ignored.

Fixes ERL-278.

Two tests are added, primarily aimed at the range splitters.

 * test_float_spills, which exercises the rare case of high floating
   point register pressure, including spill slot move coalescing.
 * test_infinite_loops, which tests that various infinite loops are
   properly compiled and do contain reduction tests (otherwise they
   would permanently hog their scheduler and not notice being sent an
   exit signal).

 * Omit bounds check in more cases.
   A test case that needs this change to omit bounds check is added.
 * Improve code generation by reformulating bounds check to decrease
   register pressure.