ash: fix quadratic matching slowdown is ${v/*foo*/repl} (really bad one)

ash: fix quadratic matching slowdown is ${v/*foo*/repl} (really bad one)

It is especially bad with patterns starting with "*".

With ASH_OPTIMIZE_FOR_SIZE=y, only those are optimized, +few bytes:
   text    data     bss     dec     hex filename
 836337     441    7564  844342   ce236 busybox_old
 836341     441    7564  844346   ce23a busybox_unstripped

With ASH_OPTIMIZE_FOR_SIZE off, we also optimize patterns _ending_ with "*",
which costs about 80 bytes:
   text    data     bss     dec     hex filename
 836656     441    7564  844661   ce375 busybox_old
 836732     441    7564  844737   ce3c1 busybox_unstripped

Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>