[Pallas] Ordered carry store for jagged row tiles

[thcmbs]

Stacked PRs:

• [Pallas] Test jagged carry with dynamic row counts #2722
• ->[Pallas] Ordered carry store for jagged row tiles #2719
• [Pallas] Implement is_row_map_axis legality gate for jagged carry #2718
• [Pallas] Jagged row-tile building blocks #2717

---

We implement the carry in this PR. When a group's first or last block isn't aligned, its window overlaps the neighbour on one row. A grid-level scratch stitches it: fold the previous group's partial into the first block (out[:S] += carry), save this group's partial from the last block (carry = out[-S:]). Saving the folded value keeps it cumulative, so several tiny groups in one row compose. Relies on groups running in order, which the grid guarantees.

Does: jagged matmul correct on f32 and bf16 across the tested layouts (shared boundaries, tiny/odd groups, block bigger than a group, multiple column tiles), checked against a per-group reference. A plain map-axis store (out[st] = 2 * jagged[st]) uses the same path.

Gaps (I left TODOs to keep track of them and iterate with follow ups):

• block larger than the whole tensor crashes (only L >= block tested);
• assumes contiguous offsets; a gap folds a stale carry (the one silent-wrong case);
• group loop must be the only grid dim (a second is refused);
• bf16 reduction over the row unsupported: not our shape, defers to [pallas] enable emit_pipeline BlockSpecs for jagged tile ranges #2716's f32-only path, can't compile. xfail tracks it; the f32 reduction already works via [pallas] enable emit_pipeline BlockSpecs for jagged tile ranges #2716.