Recover slice extents of the form `x[i:i+n]` as concrete dimensions

[khatchad]

A slice of the form x[start:start + n] has extent n on the sliced axis whenever n is statically known, regardless of start; the analysis currently degrades such slices to an unknown extent.

Witness chain (MusicTransformer-tensorflow2.0): data.py:85-93 conditionally reassigns the loaded event sequence through data = data[start:start + max_length], and Data.batch collects the per-file sequences and batches them via np.array(batch_data) (data.py:36), so the batched result's runtime shape is (batch_size, length) with length flowing from the driver. Recovering the slice extent (through the existing constant-folding and argument-resolution machinery for n) and composing the container's element shapes through np.array would tighten the batch shape from unknown toward (batch, n).

The dim-arithmetic classification follows #721: an n that resolves to a fixed integer yields a NumericDim; a fixed-but-unresolvable n yields UnresolvedDim; only None-evidence yields DynamicDim.