Range-ops uses wi_fold (implemented by each opcode)  to individually 
fold subranges one at a time and then combines them. This patch first 
calls wi_fold_in_parts which checks if one of the subranges is small, 
and if so, further splits that subrange into constants.

Currently, if a subrange is 4 or less values, then we call it 
individually for each of the 4 values. 4 was chosen as a reasonable 
tradeoff between excess work vs benefit.     We could consider 
increasing that number for -O3 perhaps.

This allows us under some circumstances to generate much more precise 
values. For instance:

     tmp_5 = x_4(D) != 0;
     _1 = (int) tmp_5;
     _2 = 255 >> _1;
     _3 = (unsigned char) _2;
     _6 = _3 * 2;
     return _6;

_1 : int [0, 1]
_2 : int [127, 255]
_3 : unsigned char [127, +INF]

we currently produce a range of [127,255] for _2.   the RHS of the shift 
(_1)  only has 2 values, so by further splitting that subrange in [0,0] 
and [1,1], we can produce a more precise range, and do better:

     y_5 = x_4(D) != 0;
     _1 = (int) y_5;
     _2 = 255 >> _1;
     _3 = (unsigned char) _2;
     return 254;

_1 : int [0, 1]
_2 : int [127, 127][255, 255]
_3 : unsigned char [127, 127][+INF, +INF]

Now _2 is a perfect 127 or 255, and we can fold this function to just 
the return value now.

This will work for many different opcodes without having to go and 
rework range-ops for them.

Bootstraps on x86_64-pc-linux-gnu  with no regressions and fixes all 3 
testcases in the PR.  pushed.

Andrew