Add TorchLogsumexpVisitor

[kit1980]

Suggest using torch.logsumexp(x) instead of torch.log(torch.sum(torch.exp(x)).

https://pytorch.org/docs/stable/generated/torch.logsumexp.html

[clee2000]

Does the README need to be updated with this new rule?

[kit1980]

Does the README need to be updated with this new rule?

Eventually, not everything is listed anyway. It should be probably generated from the code.

[malfet]

Don't you want to provide a bit more anti-patterns to make sure you don't assume that torch.log always have a two tested functions?

Suggested change

	y = torch.log(torch.sum(torch.exp(x), 1, keepdim=True) + 2.5)
	y = torch.log(torch.sum(torch.exp(x) + 2.5, 1))
	y = torch.log(torch.sum(torch.exp(x), 1, keepdim=True) + 2.5)
	y = torch.log(torch.sum(torch.exp(x) + 2.5, 1))
	y = torch.log(1 + x)

[kit1980]

Added some more tests. (Not exactly torch.log(1 + x), as it triggers another rule for log1p).

[malfet]

A bit unrelated, is there already a rule that say torch.log(torch.exp(x)) should be transformed to x, torch.sqrt(torch.sqr(x)) into torch.abs(x) and so on?

[kit1980]

A bit unrelated, is there already a rule that say torch.log(torch.exp(x)) should be transformed to x, torch.sqrt(torch.sqr(x)) into torch.abs(x) and so on?

No, is it common? Should be easy to add the checkers.