Minimise the size of padname + string buffer

Minimise the size of padname + string buffer

If we define the struct a little differently, we can begin the string
buffer two bytes into a pointer, rather than pointer-aligned.  In case
some platforms can compare pointer-aligned string faster, I added a
#define to allow that.  But on 64-bit darwin the speed is identical
either way:

$ time ./miniperl -e 'eval q|my$a;|x50000 . q|eval q<my $A>|'

I ran this three times in each mode, and the average of the user times
differed by less than 1%.