Improve the performance of long-to-double conversions on ARM.

Use a VMLA instruction instead of VADD & VMUL instructions
in long-to-double conversions on ARM.  This change reduces
code size and improves execution times (but does not alter
precision).  It trades one temporary FPU register for two
temporary core registers.

Change-Id: I1dc35bef6c12be8f305e5b46da98c2421686b60d