Cache applied with torch.compile

[Binary2355]

This PR:

• Add the Performance of Teacache and FBCache on 1xH20
• Add the Performance of Teacache and FBCache on 1xH20 and 4xH20 with torch.compile
• refracture the code to make torch.compile as efficient as possible
• Fix the bug, postpone the logic to update modulated_inputs of FBCache (which is equivalent to the paraAttn)

The performance table is listed below:

Method	Latency (s)
	without torch.compile		with torch.compile
	4xH20	1xH20	4xH20	1xH20
Baseline	2.02s	6.10s	1.81s	5.02s
use_teacache	1.60s	4.67s	1.50s	3.92s
use_fbcache	0.93s	2.51s	0.85s	2.09s

[feifeibear]

你现在遇到的问题是在_convert_transformer_backbone 先 compile 的 transformers formward。然后你apply_cache_on_transformer又改了逻辑，所以 compile 失效。

apply_cache_on_transformer 放在 _convert_transformer_backbone 函数中吧。
别在 example 中操作了。

[Binary2355]

我在_convert_transformer_backbone里进行torch compile之前保存了original_transformer, 所以apply_cache_on_transformer之后的torch compile是针对原始的transformer做的操作