Floating Point regression in NoFib's `imaginary/integrate` benchmark

It appears that recent master commits fail to pass NoFib. Specifically, imaginary/integrate fails with a stdout mismatch in modes norm and slow (but not fast, which was used by CI):

../../runstdtest/runstdtest ./integrate  -o1 integrate.stdout -o1 integrate.stdout  -ghc-timing +RTS -V0 -RTS    1000000;
0.64user 0.00system 0:00.65elapsed 100%CPU (0avgtext+0avgdata 11252maxresident)k                                                                                                                                                                                             0inputs+16outputs (0major+1491minor)pagefaults 0swaps                                                                                                                                                                                                                        
././integrate +RTS -V0 -RTS 1000000 < /dev/null                                                                                                                                                                                                                              
expected stdout not matched by reality                                                                                                                                                                                                                                       
--- integrate.stdout    2019-11-20 16:38:31.157747042 +0100                                                                                                                                                                                                                  
+++ /tmp/runtest553570.1        2021-03-20 18:12:57.094954877 +0100                                                                                                                                                                                                          
@@ -1 +1 @@                                                                                                                                                                                                                                                                  
-9.093955583391733e28                                                                                                                                                                                                                                                        
+1.9105614312239304e29
make: *** [../../mk/target.mk:101: runtests] Error 1

The first commit with which I could reproduce the issue was 4196969c. Its direct parent, f78f001c, doesn't report an output mismatch.

I suspect that the different literal desugaring produces floating point literals that are associated slightly differently.

I think that benchmark may be flawed anyway: I accumulates an ever-growing FP error. While fast still manages to output 0 at the end, the other modes report absurdly high deviations from the expected 0.

We should diagnose if we got something wrong and fix it before we release 9.2.