Port improvement to Math.random to ARM.

Port improvement to Math.random to ARM. According to issue #670,
Math.random was sped up on ia32 and x64 by creating a generated
function RandomHeapNumber, whose body is inline assembly code
that allocates a heap number, calls V8::Random to get a random
uint32, and puts the uint32 divided by 2^32 into the heapnumber.

BUG=none
TEST=none

[Erik Corry, 2010-04-22 16:59:51]

This LGTM if you update the AUTHORS file too.  Extra points for putting it back
in alphabetical order!

[rodolph, 2010-04-22 17:11:06]

On 2010/04/22 16:59:51, Erik Corry wrote:
> This LGTM if you update the AUTHORS file too.  Extra points for putting it
back
> in alphabetical order!

Authors file updated/reordered.

[Erik Corry, 2010-04-22 17:26:32]

Committed as r4477.  Thanks.

[Lasse Reichstein, 2010-04-23 12:12:50]

Complaining after the fact is still fun.

http://codereview.chromium.org/1689007/diff/5001/6001
File src/arm/codegen-arm.cc (right):

http://codereview.chromium.org/1689007/diff/5001/6001#newcode4244
src/arm/codegen-arm.cc:4244: __ CallRuntime(Runtime::kNumberUnaryMinus, 1);
It seems fragile to use a function meant for something else to create a
HeapNumber.
E.g., it would make perfect sense for someone to special-case negation of zero
in Runtime_NumberUnaryMinus (since we have a -0 constant already) - and there is
nothing there that suggests that it would break this code. 
In fact it would probably even be a good idea: The function is only called for
the smis 0 and Smi::kMinValue, or if we can't allocate a heap number in
NewSpace.

We should create an "AllocateHeapNumber" runtime function instead, if that is
what we need.