Roslyn workspace should share SyntaxTrees across project targets when possible.

From talking with @jasonmalinowski we believe there would be a large memory savings here for solutions with projects that multi-target (e.g. netstandard2.0, netcoreapp3.1, etc.). The cases that should be very safe to share are ones where the tree does not contain pp-directives, and where the lang-version for different targets are the same.

The reason for this is that pp-directives can parse differently depending on what PP directives are defined. And lang-version can affect how certain constructs are parsed (like records), as well as affect what errors are reported (like if certain constructs are or aren’t supported).

1. It would be valuable for us to collect some telemetry to see what percentage of projects that are using multi-targetting end up using the same lang-versions. On top of that, it would be good to keep track of what percentage of their files do/do-not use pp-directives.
2. we currently do keep track of stuff like pp-directives for fast-switching of debug/release. However, that tracking could be finer-grained. I believe we check if there are any directives, where we really only care about #if directives. e.g. #region and #nullable shouldn’t change things here.
3. Roslyn would be a great test bed for this as we have many targets, and tons of files. We could get a huge win here, that would translate out to projects similar to ours.
4. We may want to drive compiler APIs to assist us here. For example, it would be great if we could cheaply just ask if a file contains pp-directives. This could be a weak heuristic (potentially even at the source-text level). For example, just asking if any line starts with spaces-#-if. Or just allowing this to be determined from a green tree, etc.

[CyrusNajmabadi]

Done with https://github.com/dotnet/roslyn/pull/66003.

[jasonmalinowski]

@jaredpar We could totally aggregate that data and send it up if you wanted the telemetry, so that’d be another good argument for it.

@CyrusNajmabadi For our syntax tree indicies, I know you made changes so they could be reused across debug/release switches – do we also try to share them across multiple targets? Or do we hash the parse options into that?