`psis!` is not correct when `log_weights=true`

[ysfoo]

When log_weights=true, the function psis! only exponentiates the tail, and does not deal with the rest of the log weights correctly. For example, the cutoff passed to _psis_smooth_tail! is still in log space, which results in a DomainError in gpd_fit.

I encountered this issue when applying PSIS to log weights I obtained from a custom importance sampling problem (not in the context of PSIS-LOO).

Also, the args and kwargs in this line are not passed correctly to psis! (currently, kwargs are passed as positional arguments to psis!).