Perf improvements

[manodasanW]

This PR makes several changes to improve the perf of CsWinRT projections based on what is seen in profiling:

• Avoid constructing an ObjectReference when doing a release
• Make all the MarshalGeneric reflection lambdas lazy as we do not use all the helper types in every scenario but rather a couple in most cases. This reduces JIT cost too.
• Make all the concurrent dictionaries used as part of IDIC support lazy initialized as we don't use the caches in all scenarios and we are seeing a large # of empty ones on the finalizer.
• Fix issue where activation factories were initialized multiple times for factory and statics when we needed only one instance
• Remove creating lazy interface dictionary when there are no members and setting size in constructor

Fixes #987
Fixes #988
Fixes #825
Fixes #1000