Now the generated code doesn't need to check for end group tags, as it will skip whole groups at a time.
Currently it will ignore extraneous end group tags, which may or may not be a good thing.
Renamed ConsumeLastField to SkipLastField as it felt more natural.
Removed WireFormat.IsEndGroupTag as it's no longer useful.

This mostly fixes issue 688.

(Generated code changes coming in next commit.)

Change ReadTag and PeekTag to just use 0 as a return value for "end of stream", rather than using an awkward out parameter.
This simplifies quite a lot of code.

Generated code in next commit.

Rename ThrowHelper to Preconditions and make it public - we'll want to use it from the generated code soon.

Additionally, change it to return the value passed, and make it generic with a class constraint.
A separate method doesn't have the class constraint, for more unusual scenarios.

This requires .NET 4.5, and there are a few compatibility changes required around reflection.
Creating a PR from this to see how our CI systems handle it. Will want to add more documentation,
validation and probably tests before merging.

This is in aid of issue #590.

Changes in brief:
1. Descriptor is now the entry point for all reflection.
2. IReflectedMessage has gone; there's now a Descriptor property in IMessage, which is explicitly implemented (due to the static property).
3. FieldAccessorTable has gone away
4. IFieldAccessor and OneofFieldAccessor still exist; we *could* put the functionality straight into FieldDescriptor and OneofDescriptor... I'm unsure about that.
5. There's a temporary property MessageDescriptor.FieldAccessorsByFieldNumber to make the test changes small - we probably want this to go away
6. Discovery for delegates is now via attributes applied to properties and the Clear method of a oneof

I'm happy with 1-3.
4 I'm unsure about - feedback welcome.
5 will go away
6 I'm unsure about, both in design and implementation. Should we have a ProtobufMessageAttribute too? Should we find all the relevant attributes in MessageDescriptor and pass them down, to avoid an O(N^2) scenario?

Generated code changes coming in the next commit.

We'll see what I've missed when CI fails...