[turbofan] Make fewer assumptions about children of a serialized object

1) Strengthen the ObjectData::As* cast methods to check that the kind
is kSerializedHeapObject, because otherwise the data object is not
a subclass instance and the cast is invalid.

2) Fix errors revealed by (1) and pave way for moving away from
serialization. These changes are mechanical except for a needed
refactoring of ContextRef::previous.

Details regarding (2):

Change (1) revealed a large number of places where we incorrectly casted
object data. This went unnoticed so far because in the end we accessed
the object through the corresponding ObjectRef interface which did the
right thing depending on the data kind.

These bugs were introduced when kUnserializedReadOnlyHeapObject was
added, but they also affect the new kNeverSerializedHeapObject and would
become show stoppers as we move more objects to the latter kind. The CL
fixes all the issues that I found except one:

There's still one place left where we assume a particular subclass
instance for now (marked with a TODO). This is not a bug at the moment
but will cause CHECK failures once we move the corresponding object type
to never-serialized. A rewrite of map serialization might be needed to
resolve that.

Note:
With the changes in (2) we lose some type safety in the implementation
of the *Data classes. With some extra work that could be avoided.
However, I think it's not worth it because (i) these classes are
expected to be removed (and in the meantime to not change much), and
(ii) their wrapper *Ref classes still ensure type safety.

Bug: v8:7790
Change-Id: I9a5d03fa2f61e03c9c0ab4ac7f9869603d5be1d9
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2398537Reviewed-by: Santiago Aboy Solanes <solanes@chromium.org>
Commit-Queue: Santiago Aboy Solanes <solanes@chromium.org>
Auto-Submit: Georg Neis <neis@chromium.org>
Cr-Commit-Position: refs/heads/master@{#69775}