[OOB] models missing sdpa fusion in torch compile inference

Some OOB models are missing sdpa fusion in torch.compile, we should investigate because eager mode has this fusion

Models with SDPA ops that are decomposed (BMM+Softmax+BMM) instead of fused (flash attention). Decomposition adds overhead.

**25 model(s)** with decomposed SDPA (out of 49 models with SDPA ops).

| # | Model | SDPA Ops | B580 Status | 4080 Status | G31 Status | Correction (ms) |
|---|-------|--------:|----------|----------|----------|--------:|
| 1 | BertForQuestionAnswering_bs512 | 12 | DECOMPOSED | DECOMPOSED | DECOMPOSED | 382.8 |
| 2 | RobertaForQuestionAnswering_bs512 | 12 | DECOMPOSED | DECOMPOSED | DECOMPOSED | 382.8 |
| 3 | BartForConditionalGeneration_bs32 | 36 | DECOMPOSED | DECOMPOSED | DECOMPOSED | 362.2 |
| 4 | MBartForConditionalGeneration_bs32 | 36 | DECOMPOSED | DECOMPOSED | DECOMPOSED | 362.2 |
| 5 | AlbertForMaskedLM_bs64 | 12 | DECOMPOSED | DECOMPOSED | DECOMPOSED | 255.2 |
| 6 | AlbertForQuestionAnswering_bs64 | 12 | DECOMPOSED | DECOMPOSED | DECOMPOSED | 255.2 |
| 7 | ElectraForQuestionAnswering_bs1024 | 12 | DECOMPOSED | DECOMPOSED | DECOMPOSED | 255.2 |
| 8 | PLBartForConditionalGeneration_bs32 | 18 | DECOMPOSED | DECOMPOSED | DECOMPOSED | 135.8 |
| 9 | MobileBertForQuestionAnswering_bs4096 | 24 | DECOMPOSED | DECOMPOSED | DECOMPOSED | 134.4 |
| 10 | BertForMaskedLM_bs128 | 12 | DECOMPOSED | DECOMPOSED | DECOMPOSED | 95.7 |
| 11 | CamemBert_bs128 | 12 | DECOMPOSED | DECOMPOSED | DECOMPOSED | 95.7 |
| 12 | hf_Albert_bs128 | 12 | DECOMPOSED | DECOMPOSED | DECOMPOSED | 95.7 |
| 13 | hf_Bart_bs64 | 18 | DECOMPOSED | DECOMPOSED | DECOMPOSED | 72.4 |
| 14 | hf_Bert_large_bs512 | 24 | DECOMPOSED | DECOMPOSED | DECOMPOSED | 71.6 |
| 15 | PegasusForConditionalGeneration_bs256 | 36 | DECOMPOSED | DECOMPOSED | DECOMPOSED | 54.3 |
| 16 | DistilBertForQuestionAnswering_bs2048 | 6 | DECOMPOSED | DECOMPOSED | DECOMPOSED | 53.3 |
| 17 | hf_DistilBert_bs128 | 6 | DECOMPOSED | DECOMPOSED | DECOMPOSED | 47.9 |
| 18 | BlenderbotSmallForConditionalGeneration_bs256 | 24 | DECOMPOSED | DECOMPOSED | DECOMPOSED | 34.3 |
| 19 | M2M100ForConditionalGeneration_bs128 | 36 | DECOMPOSED | DECOMPOSED | DECOMPOSED | 27.8 |
| 20 | RobertaForCausalLM_bs32 | 12 | DECOMPOSED | DECOMPOSED | DECOMPOSED | 23.9 |
| 21 | MobileBertForMaskedLM_bs512 | 24 | DECOMPOSED | DECOMPOSED | DECOMPOSED | 17.6 |
| 22 | ElectraForCausalLM_bs64 | 12 | DECOMPOSED | DECOMPOSED | DECOMPOSED | 16.0 |
| 23 | PegasusForConditionalGeneration_bs64 | 36 | DECOMPOSED | DECOMPOSED | DECOMPOSED | 14.6 |
| 24 | DistilBertForMaskedLM_bs512 | 6 | DECOMPOSED | DECOMPOSED | DECOMPOSED | 13.5 |
| 25 | DistilBertForQuestionAnswering_bs64 | 6 | DECOMPOSED | DECOMPOSED | DECOMPOSED | 1.9 |

**Summary**: 25 models decomposed on all platforms, 0 XPU-only. Total potential savings: 3262 ms.