fix #10305 nim cpp is now nan-correct at CT

[timotheecour]

• fix nim cpp treats Nan as 0.0 (during compile time) #10305 nim cpp is now nan-correct at CT fix #10305 nim cpp is now nan-correct at CT #10310 and added a test that would fail before PR
• should fix these failing tests in CI (with nim cpp)
  • tests/float/tfloatnan.nim
  • lib/pure/math.nim
  • tests/float/tfloatmod.nim

I finally figured out root cause of #10305, but it was a rather painful debug session

Here's a minimal test case to explain what's going on:

#include <stdio.h>

#define CASE_VOLATILE // otherwise const folding optimizer kicks in
int main (int argc, char *argv[]) {
  #ifdef CASE_VOLATILE
    volatile double a1 = 0.0;
  #else
    double a1 = 0.0;
  #endif
  double a2 = 0.0;
  double a3 = a1 / a2;
  int t = a3 == 0.0;
  int t2 = a3 == a3;
  int t3 = a3 != a3;
  printf("a3=%g,t=%d,t2=%d,t3=%d\n", a3, t, t2, t3);
  return 0;
}

clang++ -o /tmp/D20181107T193702 -ffast-math -O3 $timn_D/tests/nim/all/t0103.cpp
/tmp/D20181107T193702
a3=nan,t=1,t2=1,t3=0 # BUG

without  -ffast-math
a3=nan,t=0,t2=0,t3=1

so -ffast-math affects Nan correctness;
note that :

$nimc_D/config/nim.cfg contains

clang.options.speed = "-O3"

so that explains why this issue was only affecting nim cpp (since $nimc_D/compiler/extccomp.nim defined optSpeed: " -O3 -ffast-math " for both c and cpp, but c was overridden by $nimc_D/config/nim.cfg)

note

there's a tradeoff between speed and accuracy at play here; but -d:release should be correct by default; but I added a note in doc/nimc.rst regarding more aggressive options for speed, including --passC:-ffast-math

[timotheecour]

PTAL, fixed b3 unused

[timotheecour]

PTAL