arch/arm/cpu/arm_cortexa8/omap3/cache.S: make build with older tools

The push / pop instructions used in this file are available only with
more recent tool chains:

cache.S: Assembler messages:
cache.S:133: Error: bad instruction `push {r0,r1,r2,lr}'
cache.S:160: Error: bad instruction `pop {r1,r2,r3,pc}'
cache.S:164: Error: bad instruction `push {r0,r1,r2,lr}'
cache.S:191: Error: bad instruction `pop {r1,r2,r3,pc}'

Change push/pop into stmfd/ldmfd instructions to support older
versions of binutils as well.

I verified that the modified source code generates exactly the same
binary code.

Signed-off-by: Wolfgang Denk <wd@denx.de>
Cc: Sandeep Paulraj <s-paulraj@ti.com>
Cc: Tom Rix <tom@bumblecow.com>