Fix the 'off by 1' problem in dynamically resized LoRA rank

[woct0rdho]

If I understand correctly, the index returned by index_sv_cumulative and index_sv_fro is larger by 1 than it should be. For example, torch.searchsorted works like this:

>>> cumulative_sums = torch.tensor([0.9, 0.95, 1.0])
>>> print(torch.searchsorted(cumulative_sums, 0.94))
tensor(1)
>>> print(torch.searchsorted(cumulative_sums, 0.95))
tensor(1)
>>> print(torch.searchsorted(cumulative_sums, 0.96))
tensor(2)

If the user wants to keep 94% or 95% of S, then index should be 1 (and new_rank in the function rank_resize should be 2).

This problem is particularly visible when the new LoRA is rank 1 (I've seen many LoRAs on the internet that are almost rank 1). In the example above, if the user wants to keep 80% or 90% of S, then index should be 0, and new_rank should be 1.