Add a reference to MessagePack implementation

[AArnott]

This pull request includes updates to the README.md file, enhancements to the SpanDictionary class, and improvements to the equality members of the BaseClass and DerivedClass in the test types.

Documentation Updates:

• README.md: Added a new section listing known libraries based on TypeShape, including Nerdbank.MessagePack.

Code Enhancements:

• src/TypeShape.Examples/Utilities/SpanDictionary.cs: Added a new overload for the ToSpanDictionary method to support mapping with a value selector.

Test Improvements:

• tests/TypeShape.Tests/TestTypes.cs: Implemented IEquatable<T> for BaseClass and DerivedClass, and added appropriate Equals and GetHashCode methods.

[AArnott]

Incidentally, upgrading the baseline from MessagePack v2 to v3 produces this:

Serialization

Method	Mean	Error	StdDev	Ratio	Allocated	Alloc Ratio
Serialize_TypeShape	130.9 ns	0.37 ns	0.33 ns	0.95	-	NA
Serialize_Library	137.6 ns	0.23 ns	0.21 ns	1.00	-	NA

This improves serialization slightly, making TypeShape fair slightly worse by comparison.

Deserialization

Method	Mean	Error	StdDev	Ratio	Allocated	Alloc Ratio
Deserialize_TypeShape	148.1 ns	0.63 ns	0.56 ns	1.21	32 B	1.00
Deserialize_Library	122.2 ns	0.74 ns	0.61 ns	1.00	32 B	1.00

Deserialization improves significantly, making TypeShape look quite a bit worse. This is due to using a dictionary to look up property names, which the AOT formatter avoids.

[AArnott]

Serialization is somehow faster than the Ref.Emit specialized code produced by the MessagePack library. I don't know how that's possible.

I suspect it's because in this version, we call MessagePackWriter.Write(string) directly rather than including a level of indirection that allows for some other serializer to hijack that operation at runtime to serialize something else (e.g. so that a string is only serialized once).

[eiriktsarpalis]

I can see why you went for a delegate adapter, but it does result in two layers of indirection for every call, which adds up as the object graph is being traversed.

[AArnott]

Fixed.
Can you review all the places where formatters.GetOrAdd appears in my code and tell me if I'm doing it right? I don't understand why this pattern is required, so I'm not sure...

[eiriktsarpalis]

I was suggesting that you could make Formatter<T> an abstract class and have individual types and shapes inherit from that. The downside is that you need to factor those into individual classes as opposed to inlining lambda expressions, but on the other hand splitting your code into a folder of Formatters and then a keeping a small visitor that just folds them together should make things more manageable.

[AArnott]

So far, I'm not sure what I'd do with more than the two implementations I have.
Maybe I'll need more as I get to supporting non-default constructors, etc. though.

[eiriktsarpalis]

The delayed value factory is a late binding mechanism used when a recursive type is encountered. The idea is that it's creating a wrapper for box.Result which will eventually contain a reference to the final computed value once building has completed.

For this to work, evaluating box.Result needs to be delayed until an actual serialization operation is run. Here's how it's done for CBOR:

https://github.com/eiriktsarpalis/typeshape-csharp/blob/c8c2c397beb472e716b447bb460dea626811ba18/src/TypeShape.Examples/CborSerializer/Converters/DelayedCborConverter.cs#L6-L13

And here's how the object cloner does it:

https://github.com/eiriktsarpalis/typeshape-csharp/blob/ba848263d45d05975fa7cc3a0e601b1bcdca608b/src/TypeShape.Examples/Cloner/Cloner.cs#L62

[AArnott]

@neuecc this should interest you.

[eiriktsarpalis]

Would it make sense to factor the IEquatable implementation to a separate set of types so that we coverage for both cases?

[AArnott]

I added this because I had been testing round-tripping. I'm not sure what case we're not covering when this interface isn't present.
But I did just discover IsEquatable on test cases, which probably was false for this type up to this point and why no one else hit the fact that this class didn't know how to do a by-value compare of itself.
I can remove this change if you'd like. I don't need it any more.