mpf_sub bug (essentially same as reldiff bug)

[Giulio Malinverno]

If we use the same precision for all quantities:
changing
	mpf_init2(rd, 8);
to
	mpf_init2(rd, 256);

we obtain the correct result (both in C and C++ sources)

mantissa=100000000000000000000000000000000000000146936793852785938496
diff is approx -0.5 times 2^(-128)
a:      0.1e1
b:     
0.1000000000000000000000000000000000000001469367938527859384960920671527807097273e1
rd:     
-0.1469367938527859384960920671527807097273331945965109401885939632848021574318409e-38

Best regards
Giulio Malinverno

-- 
==================================================
Necesse est quod sub certo numero omnia creata comprehendantur
==================================================