FIX AdaMSS save/load reproduction by making slice_pca SVD deterministic

[AshNicolus]

What does this PR do?

Saving an AdaMSS adapter with save_pretrained and reloading it with
from_pretrained does not reproduce the model's outputs once the adapter has
been trained.

Root cause

When an AdaMSS adapter is loaded, the layer is rebuilt from the base weights
(update_layer → slice_pca → clustering_Z). This recomputes scatter_index,
the mapping that places each subspace's contribution into the correct output
dimensions. slice_pca uses torch.svd_lowrank, which draws a random projection
from the global RNG, so the SVD result — and therefore the clustering and
scatter_index — depends on the RNG state at construction time. Because that
state differs between saving and loading, the reloaded adapter rebuilds a
different scatter_index. The trained adamss_A/adamss_B weights are restored
correctly, but are then scattered to the wrong output dimensions, so the output
changes.

This stayed hidden because AdaMSS initializes B = 0: an untrained adapter is a
no-op and reloads trivially regardless of scatter_index. The discrepancy only
appears once the adapter is trained.

clustering_Z already pins KMeans(random_state=...), so deterministic
initialization is clearly intended; the only remaining source of randomness was
svd_lowrank.

Fix

Seed a forked RNG around the svd_lowrank calls in slice_pca so the
decomposition is deterministic and the reconstructed scatter_index matches the
one used at save time. torch.random.fork_rng leaves the global RNG stream
untouched.

Tests

Added TestAdamssSaveLoad::test_save_load_reproduces_output in
tests/test_adamss_asa.py: it trains an AdaMSS adapter, saves it, perturbs the
global RNG state, reloads into a fresh base model, and asserts the outputs match.
The test fails on main and passes with this change; the existing AdaMSS tests
still pass.