Adjust AVG_POOL2D FP calculation to match conformance precision

To reduce accuracy loss for low precision floating-point types,
update the AVG_POOL2D to scale by the kernel size at the
accumulation precision.

Change-Id: I3d2d0ad0bbceb1183064530ac800ff34ca04c1f2
Signed-off-by: Jeremy Johnson <jeremy.johnson@arm.com>