[copybara-service]

Accumulate AMX flash attention directly into att_out.

Follow-up to the generic and matrix accumulation kernel changes: the AMX-BF16
and AMX-INT8 kernels now use the IN/OUT att_out directly as their
online-softmax accumulator.

• Replace the blocked accumulator layout (GetAccumTilePtr /
  StoreAccumulatorsToOutput) with row-major access to att_out through its
  MatPtrT (AccumRow helper).
• AMX-INT8: accumulator rows are only accessed per row for rows < q_count, so
  it accumulates into att_out directly. Remove the C region from
  AmxInt8WorkspaceLayout, its zero fill and the final copy.
• AMX-BF16: remove the per-call heap allocation and zero fill of C. Full
  32-query blocks Tile64BLoad/Store straight from att_out with its row
  stride. AMX tiles span all 32 rows of a block, so a partial last block uses
  a MatPtrT view over a 32 x qkv_dim scratch vector (allocated only when
  there is a partial block), copied in from and back out to att_out.
• Document att_out as accumulated in place by all kernels.
• Extend the shared test setup to the AMX kernels and add
  TestTiledFlashAttentionAMXChainedCalls with 64, 40 and 15 queries (full
  blocks, full + partial, partial only), via a new query_counts parameter of
  RunTiledFlashAttentionChainingTest. It skips visibly (GTEST_SKIP) on hosts
  without AMX instead of silently passing.