rope: make the CPU reference match the device's angle-row convention

[atassis]

Fixes #188.

reference() used cos.repeat(rep, 1) to stretch an angle_rows-row LUT to rows rows, which tiles the whole block (row r uses angle row r % angle_rows). The device kernel (design.py's core_body) applies one angle row to rows / angle_rows consecutive input rows instead (row r uses angle row r // (rows / angle_rows)). The two conventions only agree at angle_rows in {1, rows}, so this was invisible until something calls reference() at 1 < angle_rows < rows -- exactly the shape llama_npu.py's prefill RoPE uses (rows=prompt_len*n_heads, angle_rows=prompt_len), though nothing currently routes through reference() at that shape. repeat_interleave is the direct fix: it repeats each row in place instead of tiling the block.

I'm recommending the device kernel as canonical over reference(), not the other way around -- see the linked issue for why.

Added

• iron/tests/operators/rope_reference_convention.py: builds a ground truth directly from the device's row-to-angle mapping and checks reference() against it across several (rows, angle_rows) shapes, including the decisive 1 < angle_rows < rows case.

Changed

• iron/operators/rope/reference.py: cos.repeat(rep, 1) / sin.repeat(rep, 1) -> repeat_interleave(rep, dim=0), and the docstring now says which convention and why.

Removed

Evidence

$ PYTHONPATH=. python3 -m pytest --noconftest iron/tests/operators/rope_reference_convention.py -v
======================== test session starts ========================
collected 2 items

rope_reference_convention.py::test_reference_matches_device_convention_for_batched_angle_rows PASSED
rope_reference_convention.py::test_reference_matches_device_convention_across_shapes PASSED

======================== 2 passed in 0.89s ========================

Before the fix, the first test's rows=6, angle_rows=3 case shows 4 of 6 rows mismatching the device-convention ground truth (see the linked issue for the exact numbers); after repeat_interleave, all shapes tested ((6,3), (8,2), (1024,1), (4,4), (13,13), (12,4)) match exactly.

Not touched: generate_golden_reference()/apply_rope(), which rope/test.py actually uses for its device-comparison golden -- that path already encodes the consecutive-row convention correctly via its .transpose(0, 1), so it isn't affected by this bug and this PR doesn't change it.

PR Merge Checklist

1. The PR is rebased on the latest devel commit and pointing to devel.
2. Your PR has been reviewed and approved.
3. All checks are passing.