Avoid clipping in short->float stereo downmixing.

Previously I'd said this would require a large stack buffer, which
 was true if you simply wanted to re-use the existing
 op_short2float_filter() and floating-point op_stereo_filter().
But the latter is not normally compiled for fixed-point anyway, so
 we can instead write new code that doesn't need the stack buffer
 without harming anything.