[TBDGen] Fix check for global accessors

[harlanhaskins]

Previously, TBDGen skipped emitting lazy initializers for globals that
appeared in any file with an entry point. This breaks, however on files
that have an NSApplicationMain/UIApplicationMain class in them, where
the entry point is synthesized but top-level globals are not locally
scoped. This change re-uses SILGen's check and only skips variable
declarations that appear at top level in a script mode file.

Resolves rdar://43549749

[harlanhaskins]

@swift-ci please smoke test

[slavapestov]

Does it ever make sense for script mode declarations to have public symbols? Maybe we should fix the linkage computation for the initializer instead

[slavapestov]

Also why would you want to generate a TBD for something with a main() function? Isn’t TBD only for dynamic libraries?

[harlanhaskins]

If -enable-testing is passed, global lazy inits get public symbols, but that's a useless flag if you're compiling in script mode, but we could detect that and keep them private.

Currently, in Debug on Apple platforms, we validate the TBD file symbol generation against generated IR unless explicitly told not to via -validate-tbd-against-ir=none, so even if we're not going to emit a TBD file directly, we still support the symbol validation logic. Maybe we could revisit that policy for modules with a main.

[jrose-apple]

It's still sensible to try to parallelize building test targets with building the app target.

[harlanhaskins]

It would probably be a good idea to revisit visibility for top-level symbols in script files, but for now I'm gonna merge this to unblock the bots.