Distinguish uninferred dimensions from rank-0 scalars

[t-kalinowski]

Variable@dims currently uses NULL to represent a rank-0 scalar. This is reflected in the class comment, the rank and is_scalar properties, and scalar Variable() construction throughout the package.

However, Variable() is also used as a placeholder while its type and shape are being inferred. These placeholders have dims = NULL, so they are treated as rank-0 scalars before their shape is known. This makes an uninferred shape indistinguishable from an inferred scalar.

I think we should use the following representation:

@dims	Meaning
NULL	shape and rank have not been inferred
list()	rank-0 scalar
list(1L)	rank-1 value with extent 1
list(NA_integer_)	rank-1 value with an unknown extent
list(NA_integer_, NA_integer_)	rank-2 value with unknown extents

Under this representation, declarations such as integer() would produce dims = list(). Unknown extents of a known-rank value would continue to use NA per axis.

The implementation will require:

• preserving the distinction between NULL and list() in the dims setter;
• making scalar Variable() construction pass dims = list() explicitly;
• replacing other deliberate uses of NULL for scalar results;
• making shape-sensitive operations fail clearly when dimensions have not been inferred;
• auditing @rank, @is_scalar, passes_as_scalar(), and is.null(@dims) consumers;
• adding public tests that exercise rank-0 declarations, generated code, runtime results, and shape inference.

This is a follow-up to #137. The setter change there remains consistent with the current representation; this issue proposes changing that representation separately.