* [Standalone] Decouple MKLDNN primitive build from code generation

This patch introduces a new pass, MKLDNNPrimitiveBuildPass, which
iterates over all the ops assigned to MKLDNN and builds their
corresponding primitives. Primitive indexes are stored in MKLDNNEmitter
and can easily be retrieved with the get_primitive_index(node)
interface. This decouples the creation of primitives from codegen and
fixes the problem of MKLDNN primitives being created twice
(CommonFunctionCollection pass and codegen).

Current assertions only allow the creation of a single primitive per
node but it should be simple to remove this when needed. Using a pass
might not be the best approach here but I found it convenient for the
current needs and it should be straightforward to convert into a utility,
if needed.

These changes caused a conflict with recently introduced
'build_quantized_inner_product*'. These new build methods will be ported
in a follow up patch to new build approach.

* Remove unrelated comment

* Remove TensorView code

* Set m_node_primitive_map from MKLDNNPrimitiveBuildPass

* Move node->primitive map from mkldnn pass to external function

* Fix struct/class inconsistency en fw declaration