fix(test): attend_batched mask test fails on ROCm on a fully masked row #2219

Problem

Since PR #2210 (KV attention dispatch moved into the cache; #2171, Phase 4a of epic #2166, implemented on a non-ROCm host), cargo test --profile test-fast --features rocm -p mlxcel-core --lib cache::attend::attend_tests on gfx1151 gives 10 passed, 1 failed: batched_masked_rows_get_their_own_mask_row (src/lib/mlxcel-core/src/cache/attend_tests.rs:213), the assert_eq!(to_vec_f32(&got), to_vec_f32(&want)) on line 252.

The batched output and the per-row reference are bit-identical where finite and both NaN at the same 256 positions, all in batch row 1. The assertion fails only because NaN != NaN, so attend_batched is correct on ROCm.

Cause

The test masks key 0 for every query of row 1 and keys s > t causally, so query t=0 of row 1 has every key masked. The f16 mask value -1e9 rounds to -inf. On ROCm an array-masked SDPA with head dim 64 has no fused kernel (supports_sdpa_vector, supports_sdpa_flash, and supports_sdpa_flash_wmma all reject array masks), so MLX's graph fallback runs add(mask) plus softmax, and softmax over an all -inf row is NaN. On Metal the same shape (qL <= 8 with an array mask) runs sdpa_vector, which skips keys whose mask is below finite_min and returns 0 for an empty row, so both sides are 0 and the test passes.

Fix

• Compare the two outputs bitwise with f32::to_bits, so NaN in the same slot counts as equal.
• Keep the mask shape and document why the fully masked query is there.
• Separately confirm whether production batched prefill on ROCm can produce a fully masked query row.

Acceptance Criteria

• cache::attend::attend_tests passes 11 of 11 with --features rocm on gfx1151
• The test still fails when attend_batched uses the wrong mask row (checked with a deliberate mutation)
• Production exposure to fully masked rows on ROCm is measured and reported (separate issue if real)

References: #2210, #2171, epic #2166, ROCm verification issue #2192.