Add new transformers model type: Bart

[wangyems]

Description: Describe your changes.
issues that apply:
#8637
#7796
-enable the a new type as "bart" in transformer optimizer
-enhance the reshape fusion for huggingface(4.9) bart model
-fuse the self-attention in encoder

Linux/V-100/intra_num_threads=24

Basically benefit cpu with smaller size. For large size one could disable attention fusion

Before:

model	EP	precision	b1_s128	b1_s512	b2_s128	b2_s512
bart-base	cpu	fp32	71.94	339.22	131.94	432.31
bart-large	cpu	fp32	213.95	656.69	377.22	1287.50
bart-base	cuda	fp32	5.30	14.67	10.89	25.53
bart-large	cuda	fp32	13.82	43.64	21.25	81.58

After:

model	EP	precision	b1_s128	b1_s512	b2_s128	b2_s512
bart-base	cpu	fp32	59.73	219.49	102.28	414.53
bart-large	cpu	fp32	189.58	699.35	328.35	1323.91
bart-base	cuda	fp32	5.12	14.61	10.00	25.43
bart-large	cuda	fp32	13.22	43.29	20.89	81.12

TODO: (Will include in the separate PR when get to them)
-accepts attention masks
-implement cross-attention in decoder
-fuse the decoder attention

Motivation and Context

• Why is this change required? What problem does it solve?
• If it fixes an open issue, please link to the issue here.

[yufenglee]

Do you perf improvement for fp16 on GPU? And why there is no improvement for fp32 on GPU?