Add MultiHeadAttention

[marko-fabo-htec]

Add support for MultiHeadAttention operator.
Description: https://github.com/microsoft/onnxruntime/blob/main/docs/ContribOperators.md#com.microsoft.MultiHeadAttention

TODOs:

• add key_padding_mask input handling
• add attention_bias input handling
• add past_key and past_value inputs handling
• add parse tests
• check performance

[codecov]

Codecov Report

Attention: Patch coverage is 77.97619% with 37 lines in your changes missing coverage. Please review.

Project coverage is 92.08%. Comparing base (da545d2) to head (9f0d026).

Files with missing lines	Patch %	Lines
src/onnx/parse_multi_head_attention.cpp	77.97%	37 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #3650      +/-   ##
===========================================
- Coverage    92.19%   92.08%   -0.11%     
===========================================
  Files          513      514       +1     
  Lines        21633    21801     +168     
===========================================
+ Hits         19945    20076     +131     
- Misses        1688     1725      +37     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[migraphx-bot]

Test	Batch	Rate new 9f0d02	Rate old da545d	Diff	Compare
torchvision-resnet50	64	3,256.22	3,255.84	0.01%	✅
torchvision-resnet50_fp16	64	6,983.33	6,991.64	-0.12%	✅
torchvision-densenet121	32	2,432.22	2,431.85	0.02%	✅
torchvision-densenet121_fp16	32	4,072.97	4,070.27	0.07%	✅
torchvision-inceptionv3	32	1,627.94	1,628.09	-0.01%	✅
torchvision-inceptionv3_fp16	32	2,746.22	2,747.87	-0.06%	✅
cadene-inceptionv4	16	764.30	764.69	-0.05%	✅
cadene-resnext64x4	16	806.98	806.89	0.01%	✅
slim-mobilenet	64	7,459.90	7,464.92	-0.07%	✅
slim-nasnetalarge	64	208.45	208.40	0.03%	✅
slim-resnet50v2	64	3,441.56	3,441.42	0.00%	✅
bert-mrpc-onnx	8	1,151.41	1,145.08	0.55%	✅
bert-mrpc-tf	1	462.81	461.72	0.24%	✅
pytorch-examples-wlang-gru	1	422.51	429.74	-1.68%	✅
pytorch-examples-wlang-lstm	1	381.78	480.37	-20.52%	🔴
torchvision-resnet50_1	1	778.70	770.17	1.11%	✅
cadene-dpn92_1	1	408.16	403.49	1.16%	✅
cadene-resnext101_1	1	382.75	381.97	0.20%	✅
onnx-taau-downsample	1	345.46	346.00	-0.16%	✅
dlrm-criteoterabyte	1	33.33	33.31	0.05%	✅
dlrm-criteoterabyte_fp16	1	52.72	52.72	0.00%	✅
agentmodel	1	8,341.74	8,212.54	1.57%	✅
unet_fp16	2	58.85	58.71	0.24%	✅
resnet50v1_fp16	1	932.36	938.46	-0.65%	✅
resnet50v1_int8	1	1,006.10	999.92	0.62%	✅
bert_base_cased_fp16	64	1,170.38	1,169.75	0.05%	✅
bert_large_uncased_fp16	32	362.84	363.16	-0.09%	✅
bert_large_fp16	1	198.52	200.11	-0.80%	✅
distilgpt2_fp16	16	2,200.60	2,198.98	0.07%	✅
yolov5s	1	526.32	533.15	-1.28%	✅
tinyllama	1	43.34	43.41	-0.17%	✅
vicuna-fastchat	1	177.16	173.49	2.11%	✅
whisper-tiny-encoder	1	417.99	417.39	0.14%	✅
whisper-tiny-decoder	1	435.90	427.99	1.85%	✅

This build is not recommended to merge 🔴

[migraphx-bot]

     ✅ bert-mrpc-onnx: PASSED: MIGraphX meets tolerance

     ✅ bert-mrpc-tf: PASSED: MIGraphX meets tolerance

     ✅ pytorch-examples-wlang-gru: PASSED: MIGraphX meets tolerance

     ✅ pytorch-examples-wlang-lstm: PASSED: MIGraphX meets tolerance

     ✅ torchvision-resnet50_1: PASSED: MIGraphX meets tolerance

     ✅ cadene-dpn92_1: PASSED: MIGraphX meets tolerance

     ✅ cadene-resnext101_1: PASSED: MIGraphX meets tolerance

     ✅ dlrm-criteoterabyte: PASSED: MIGraphX meets tolerance

     ✅ agentmodel: PASSED: MIGraphX meets tolerance

     ✅ unet: PASSED: MIGraphX meets tolerance

     ✅ resnet50v1: PASSED: MIGraphX meets tolerance

     ✅ bert_base_cased_fp16: PASSED: MIGraphX meets tolerance

🔴bert_large_uncased_fp16: FAILED: MIGraphX is not within tolerance - check verbose output

     ✅ bert_large: PASSED: MIGraphX meets tolerance

     ✅ yolov5s: PASSED: MIGraphX meets tolerance

     ✅ tinyllama: PASSED: MIGraphX meets tolerance

     ✅ vicuna-fastchat: PASSED: MIGraphX meets tolerance

     ✅ whisper-tiny-encoder: PASSED: MIGraphX meets tolerance

     ✅ whisper-tiny-decoder: PASSED: MIGraphX meets tolerance

     ✅ distilgpt2_fp16: PASSED: MIGraphX meets tolerance