Use unchecked function callbacks for better performance

[kpreisser]

Hi,

this PR implements unchecked function variants (bytecodealliance/wasmtime#3350) for the generic callback overloads added by #163, which improves performance.

This adds a ValueRaw struct that maps to wasmtime_val_raw_t and an IValueRawConverter<T> interface similar to the existing ones, which are then used by the unchecked callback functions.

Since the .NET side knows the exact parameter and result types used by the callback, this should be safe. (However I'm not very familiar with wasmtime's resource management, so I would appreciate if you can review especially the handling in the FuncRefValueRawConverter and GenericValueRawConverter.)

This also improves the NRT annotations for callback delegates (note that for <auto-generated> files, we must explicitely add #nullable enable even if the project already enables it), and fixes a regression that prevented to define callbacks using an interface as generic type parameter (because that is currently not allowed by ValueBox.Converter<T>(), but it is allowed by Value.TryGetKind()).

Additionally, callbacks that use neither Caller nor Function as parameters are now allocation-free when being called, since the Caller instance will no longer be created as it isn't needed in these cases.

When trying to run the benchmarks from #163 (but using 6 instead of 5 iterations) with .NET 6.0.11 on Windows 10 (Build 19044) x64, I get the following results:

Benchmark 1 (Action<int>):

Without this PR:

Elapsed: 00:00:00.1257743
Elapsed: 00:00:00.1165437
Elapsed: 00:00:00.1098291
Elapsed: 00:00:00.1045918
Elapsed: 00:00:00.1042576
Elapsed: 00:00:00.1047506

With this PR:

Elapsed: 00:00:00.0524017
Elapsed: 00:00:00.0489112
Elapsed: 00:00:00.0494636
Elapsed: 00:00:00.0482166
Elapsed: 00:00:00.0423121
Elapsed: 00:00:00.0429405

Benchmark 2 (Action<int, float, long>):

Without this PR:

Elapsed: 00:00:00.2216621
Elapsed: 00:00:00.1976731
Elapsed: 00:00:00.1968243
Elapsed: 00:00:00.1969110
Elapsed: 00:00:00.1974007
Elapsed: 00:00:00.1961490

With this PR:

Elapsed: 00:00:00.0652740
Elapsed: 00:00:00.0631706
Elapsed: 00:00:00.0615255
Elapsed: 00:00:00.0497510
Elapsed: 00:00:00.0505233
Elapsed: 00:00:00.0497147

Benchmark 3 (Func<int, float, long, ValueTuple<int, int, long>>):

Without this PR:

Elapsed: 00:00:00.4356896
Elapsed: 00:00:00.3905873
Elapsed: 00:00:00.3896729
Elapsed: 00:00:00.3901601
Elapsed: 00:00:00.3902660
Elapsed: 00:00:00.3917198

With this PR :

Elapsed: 00:00:00.0881153
Elapsed: 00:00:00.0869675
Elapsed: 00:00:00.0849886
Elapsed: 00:00:00.0742805
Elapsed: 00:00:00.0611120
Elapsed: 00:00:00.0619580

With the last benchmark, this seems to be roughly a 6x improvement comparing to the current state (without this PR).

What do you think?

Thanks!

[peterhuene]

@kpreisser thanks for all this work! I paid close attention to the externref/funcref raw value conversion and it seems correct to me.