Fix ALiBi distance matrix to use a proper (q, k) outer difference

[systemblueio]

Proposed changes

ALiBi.alibiMatrix builds the relative-distance matrix like this:

let x1 = MLXArray(key.offset ..< key.qSequenceLength).expandedDimensions(axis: 1)
let x2 = MLXArray(0 ..< key.kSequenceLength).expandedDimensions(axis: 1)
let distanceMatrix = -abs(expandedDimensions((x1 - x2), axes: [0, 1]))

Both vectors are expanded on axis 1, so the shapes are (q, 1) and (k, 1) and x1 - x2 is (q, 1) - (k, 1). That broadcast throws whenever the query and key lengths differ, and when they are equal it collapses to a constant column ((offset + i) - i) instead of the (i - j) outer difference, so the ALiBi bias is wrong in the cases it is meant to handle.

This expands the key vector on axis 0 instead, so the difference is (q, 1) - (1, k) and broadcasts to the intended (q, k) matrix, matching the reference Python implementation in mlx (x1[:, None] - x2[None, :]). Adds MLXNNPositionalEncodingTests covering the relative-distance values and a differing query/key length case. There is no existing issue for this; the bug was found while reading the layer.

Note: I verified swift build --build-tests and swift format lint --strict locally, but could not run the suite here because SwiftPM on the command line cannot build the Metal shaders (per the README), so the new tests are scoped to the CPU stream and exercise on CI.

Checklist

Put an x in the boxes that apply.

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)