repeat: split cols against every constraint the BD imposes, and add test coverage

[atassis]

Stacked on #157: the new tests gate at exact equality, which #157 is what makes
expressible. Its two commits are in this diff and drop out when it merges.

repeat has no tests, and could not have had: rows=8, cols=512, repeat=4 does not
build.

Both taps split cols as

sizes=[repeat, rows, cols // cols_split, cols_split]   strides=[..., cols_split, 1]

so the innermost dim counts chunks and the contiguous run sits one dim out. cols_split
is 1 whenever no split is needed, which makes the innermost dim one element. At bf16
that is two bytes, not a whole 32-bit word, and verifyStridesWraps rejects it:

'aie.dma_bd' op Transfer sizes must be multiples of 4 bytes. 1 elements at 2 bytes
each equal 2 bytes, which is not divisible by 4.
aie.dma_bd(%arg0 : memref<8x512xbf16> ... sizes = [4, 8, 512, 1] strides = [0, 512, 1, 1])

The 512-element stride-1 run the transfer is actually made of is right there in dim 2.

The reason this survived: llama_npu.py passes cols = prompt_len * head_dim, far above
1023, so it lands at cols_split = 256 and its innermost dim is a healthy 256 elements.
The only caller in the tree never trips it; every small-cols shape does, which is what a
test would use.

Added

• iron/operators/repeat/test.py. Five device arms at rel_tol=abs_tol=0 (repeat moves
  data and computes nothing) spanning cols_split 1, 2, 4 and 256 -- the last being the
  shape llama_npu.py actually dispatches -- plus three arms pinning the shapes that have
  no legal split. 8/8.

Changed

• iron/operators/repeat/design.py: swap the two inner dims so the innermost carries
  cols // cols_split at stride 1. The address sequence is byte-identical either way --
  the nested (outer, inner) walk of a linear range is the same walk regardless of which
  factor is outer -- so the change is only in which dim the hardware sees as innermost.

  The swap moves cols_split onto the next dim out, which carries the same 10-bit wrap
  field, so bounding the chunk length alone is not sufficient. The selection loop now
  searches for a divisor satisfying all three constraints at once: chunk <= 1023, chunk a
  whole number of 32-bit words, and count <= 1023. Where no divisor satisfies them -- an
  odd cols, a prime above 1023, or twice such a prime -- it raises rather than emitting a
  tap the BD verifier rejects less legibly. Every cols that worked before selects the
  same cols_split, including llama's 131072.

Removed

Evidence

8/8 on device, run without the not extensive filter so both extensive arms -- including
the llama shape -- actually executed. The emitted BD for the llama arm is
sizes = [4, 8, 256, 512] strides = [0, 131072, 512, 1].