Update inexact round bounds and computation

The current version of the specification only allows output zero points
to be subtracted as integers, so the FP32 implementation was updated to
reflect that.

The error bound was fixed to now use the output before the zero point is
added to compute the error scale.

Signed-off-by: Ian Tayler Lessa <ian.taylerlessa@arm.com>
Change-Id: I5d2834b22180137880ffa905567ff1722d61c060