[runtime] upgrade libtorch version to 2.1.0

[xingchensong]

try to fix #2417 && #2393

TODO

• pass compilation
  • linux cpu
  • linux gpu
  • windows
  • macos
  • android libtorch (next PR)
• wer verification
• speed comparison with 1.13.0

libtorch	aishell-1 test (chunk 16 left -1, ctc greedy search)	rtf
2.1.0+cpu	5.87 % N=104765 C=98725 S=5892 D=148 I=107	0.04972 (int8)
2.1.0+gpu	5.80 % N=104765 C=98786 S=5833 D=146 I=101	0.07579 (fp32)
1.13.0+cpu	5.85 % N=104765 C=98746 S=5871 D=148 I=110	0.05211 (int8)

[xingchensong]

why c++17?

[Mddct]

https://github.com/wenet-e2e/wenet/blob/main/wenet/transformer/encoder.py#L265

记得这里 att mask to bias， sdpa的bool的mask 算的不太对（batch下不太对， 单条的不确定要不要to bias），

[xingchensong]

https://github.com/wenet-e2e/wenet/blob/main/wenet/transformer/encoder.py#L265

记得这里 att mask to bias， sdpa的bool的mask 算的不太对（batch下不太对， 单条的不确定要不要to bias），

jit下用不到attention mask，哪个一直都是 torch.empty((0, 0, 0))，因为cache和current chunk都是valid

[xingchensong]

Why have we chosen libtorch 2.1.0 over the 2.1.2 version, which is in strict alignment with the PyTorch version?

As of March 19, 2024, libtorch 2.1.0 represents the latest available release for download from the official repository at https://download.pytorch.org/libtorch/cpu