Minor SDPA optimizations

[cglagovichTT]

Ticket

Subtask of #16557

Problem description

SDPA has quite a few unnecessary operations which make it inefficient, especially as sequence length grows.

What's changed

• Remove all block copies by efficiently ping-ponging buffers with aliases and std::swap
• Use L1 accumulation to update intermediate output to avoid an extra unpack/add/pack
• Reuse DST in mul_block_bcast_cols_accumulate
• Fix bug with DST reuse and dhead=96
• Don't allocate CB for mask if a mask isn't used. This reduces memory waste and enables larger chunk sizes

For the following test case, we get a nice 1.084x speedup.
tests/tt_eager/python_api_testing/unit_testing/misc/test_scaled_dot_product_attention.py::test_sdpa_tt_large_seq[1-8-1-131072-128-k128-q128-bf16]

ID  Total %  Bound  OP Code                    Device Time   Op-to-Op Gap  Cores  DRAM  DRAM %  FLOPs  FLOPs %  Math Fidelity
----------------------------------------------------------------------------------------------------------------------------------
2  100.0 %         OLD ScaledDotProductAttention  1,949,831 us                   64                                BF16, BF16 => BF16
2  100.0 %         NEW ScaledDotProductAttention  1,798,710 us                   64                                BF16, BF16 => BF16

Checklist

• Post commit CI passes https://github.com/tenstorrent/tt-metal/actions/runs/12831544820
• Model regression CI testing passes https://github.com/tenstorrent/tt-metal/actions/runs/12831557373

[caixunshiren]

Overall looks good!

[caixunshiren]

Maybe better error messaging?

[cglagovichTT]

I found that one of the optimizations in this branch, using mul_block_bcast_cols to write directly to cb_out, leads to inexplicable PCC issues in Llama tests. I was able to reproduce this in a chunked prefill unit test, but it's unclear why this optimization leads to different outputs from before.

[caixunshiren]

LGTM