Don't inline classes in constructor if possible

[t-kalinowski]

This PR changes a constructor default formal value from an inlined class call, to a call like pkgname::classname(), if class@package is not NULL.

closes #477

With this PR:

library(S7)
Coordinate <- new_class(
  "Coordinate",
  properties = list(
    x = class_double,
    y = class_double
  ),
  package = "draw"
)

Shape <- new_class(
  "Shape",
  properties = list(
    position = Coordinate, # would normally inline Coordinate
    name = class_character
  ),
  package = "draw"
)

str(formals(Shape))
#> Dotted pair list of 2
#>  $ position: language draw::Coordinate()
#>  $ name    : chr(0)
stopifnot(identical(
  formals(Shape)$position,
  quote(draw::Coordinate())
))

[t-kalinowski]

This approach of using pkgname::classname() might not work, because new_class() is expected to run at package build time, before package exports are fully known (i.e., before a call like :: or getNamespaceExports() or getExportedValue() will work correctly).

We might have to change the parent environment of the constructor to the environment where new_class() is called. Though, that also brings challenges even in S7 itself, like:

** testing if installed package can be loaded from temporary location
Error: package or namespace load failed for ‘S7’:
 .onLoad failed in loadNamespace() for 'S7', details:
  call: NULL
  error: 'new_object' is not an exported object from 'namespace:S7'
Error: loading failed

[lawremi]

Qualifying calls by the namespace should only be necessary when referencing class from another package. Would avoiding unnecessary qualification resolve the issue?

[hadley]

This is starting to feel related to #341

[t-kalinowski]

OOPWG notes: This should be fixed and merged before release.