DMMHA: add unit tests; fix CPU, CUDA kernel (microsoft#22567)

### Description

Fixes:
(1) cpu kernel: applying scale before bias and mask like other MHA ops
(2) cpu kernel: correct offset during appending past to present.
(3) cuda kernel: apply mask if provided; fix output_qk offset.

Add DMMHA unit tests

onnxruntime/test/contrib_ops/decoder_masked_multihead_attention_op_test.cc

@@ -757,7 +757,7 @@ static void TestDecoderMaskedMultiHeadAttention(bool is_cross_attn = true, bool
 
   OpTester tester("DecoderMaskedMultiHeadAttention", 1, onnxruntime::kMSDomain);
   FixedPatternValueGenerator generator{};
-  RandomValueGenerator random{123};
+  RandomValueGenerator random{};
 
   // Attributes
   tester.AddAttribute<int64_t>("num_heads", static_cast<int64_t>(num_heads));