Make trace() work for functions whose S3 class has multiple strings

[hadley]

trace() failed for any function with a multi-string S3 class attribute, e.g. S7 generics and methods (class c("S7_method", "function", "S7_object")), with 'length = 3' in coercion to 'logical(1)' because the class vector was passed where a single class name was expected. Reported via RConsortium/S7#584.

In methods:::.TraceWithMethods:

• Use .class1() rather than class() to derive the "<class>WithTrace" class name, so the trace class extends the (S4 view of) the most specific class instead of setClass() receiving a vector of names.

In methods:::.initTraceable:

• Use .class1() likewise for the old class lookup.

• Skip the as(.Object, oldClass) <- def slot copy when oldClass is virtual, i.e. an S3 class registered by setOldClass(). Such classes have no slots to copy beyond .Data, and for them the coercion replaced .Object wholesale with def, losing the trace wrapper.

• Extract the data part with slot(def, ".Data") instead of def@.Data: since R 4.3, @ on non-S4 objects dispatches on S3 methods, so a package method such as S7's @.S7_object intercepted the call and errored.

Tracing an S7 method now works end to end with no changes to S7.

Tracing an S7 generic additionally needs S7's C-level change (RConsortium/S7#695): the traced binding is an S4 object of class "S7_genericWithTrace" whose link to "S7_generic" exists only in the S4 class table, so the Rf_inherits(generic, "S7_generic") check in S7's method-dispatch.c cannot see it and S7 must unwrap the "original" attribute itself. That is inherent to how trace() wraps functions and not fixable in R: an S4 object's class attribute is its S4 class, and making inherits() consult S4 inheritance would be far more invasive. S7's R-level workaround (pre-registering the WithTrace classes) remains useful on R versions without this fix and for property access and printing of traced generics.

Adds a regression test in tests/classes-methods.R and a NEWS entry.