MLCE-1138 Issue with Delegate supporting FP16 models

 * Fixed issue where backends were asked to support FP16 layers that would
   be optimized out.
 * Fixed issue where backends were asked to support non-constant filter
   and bias tensors when those tensors would be replaced by constant
   tensors during optimization.

Signed-off-by: Mike Kelly <mike.kelly@arm.com>
Change-Id: Ib54b9cb99d5014e27172841a665daf57d1d5b23d