Add UnsafeAccessor support for generic types

[AndriySvyryd]

• Group UnsafeAccessor methods by target type
• Generate fully qualified lambda signatures
• Generate shorter UnsafeAccessor names
• Reenable test code that uses the compiled model
• Don't use Roslyn Formatter or Simplifier to make generation faster and avoid artifacts

Fixes #33773
Fixes #33470

[roji]

@AndriySvyryd re removing Roslyn Formatter/Simplifier, can we first verify that these represent a significant part of the slowdown before removing them (/cc @maumar)? We of course don't have to produce nice code, but it's definitely a better if we can...

[roji]

Another question: I don't remember any more where we landed, but do we actually use the unsafe accessors generated by LinqToCSharpTranslator? IIRC we've got the unsafe accessors you generate in the compiled model, ideally we'd just use those in the interceptors rather than generate them again...

[roji]

This change is great, but just asking (@maumar) - is the slowness in generating the code, or in compiling it? Because this change helps the latter rather than the former.

[maumar]

I haven't had a chance to investigate this yet. Hopefully within next few days

[AndriySvyryd]

@AndriySvyryd re removing Roslyn Formatter/Simplifier, can we first verify that these represent a significant part of the slowdown before removing them (/cc @maumar)? We of course don't have to produce nice code, but it's definitely a better if we can...

This change makes pre-compiled query generation for AOTSample100Queries take about 50% less time.

[AndriySvyryd]

Another question: I don't remember any more where we landed, but do we actually use the unsafe accessors generated by LinqToCSharpTranslator? IIRC we've got the unsafe accessors you generate in the compiled model, ideally we'd just use those in the interceptors rather than generate them again...

We always use the unsafe accessors from the compiled model. There's an Assert checking that LinqToCSharpTranslator doesn't produce any new accessors

[roji]

@AndriySvyryd re removing Roslyn Formatter/Simplifier, can we first verify that these represent a significant part of the slowdown before removing them (/cc @maumar)? We of course don't have to produce nice code, but it's definitely a better if we can...

This change makes pre-compiled query generation for AOTSample100Queries take about 50% less time.

... wow!

[roji]

LGTM, but see comment about Generate/TryGenerate.