Add custom logsigmoid grad for PyTorch executor

[mruberry]

See #1520, where some details of PyTorch's logsigmoid grad implementation are discussed and supported.

I think we can avoid the extra creation of the CUDA buffer tensor (even though it's empty) by implementing a thunder grad formula without it, then adding a special grad formula for logsigmoid when it's being executed by the PyTorch executor that has the buffer tensor.

cc @apaz-cli