Add asymmetric head_size regression test for MEA fallback
Add TestONNXAttentionMHAAsymmetricHeadSize to verify that MEA gracefully
falls back to unfused attention when head_size != v_head_size with past_key
present (decode phase). Without the eligibility guard in ComputeInternal,
this would crash with ORT_ENFORCE in LaunchConcatNewToPastKV.
To support this test, add v_head_size field to AttentionConfig (defaults
to 0 = same as head_size) and propagate it through the ONNX graph builder
and io_binding helpers for V-related shapes (V input, past_value,
present_value, Y output). All existing tests are unaffected since they
don't set v_head_size.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Agent-signed-off: Developer (b0ebe545) [claude-opus-4.6]
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
