Allow custom torch.compile backends to customize what gets logged to tlparse

[zou3519]

Internal x-ref

We also see this with vLLM. internal tlparse.

vLLM's custom backend invokes AOTAutograd+Inductor multiple times (see image), some way to customize the behavior of that, like tag each AOTAutograd+Inductor run, would make things a bit easier to grok.