[BUG] `make_regression` is not deterministic with `shuffle=True`.

[trivialfis]

Describe the bug
As in the title. Running it multiple times with the same seed generates different results.

Steps/Code to reproduce bug

from cuml.datasets import make_regression
import cupy as cp

X1, y1 = make_regression(
    n_samples=100, n_features=5, n_informative=5, random_state=42
)
X2, y2 = make_regression(
    n_samples=100, n_features=5, n_informative=5, random_state=42
)

assert cp.array_equal(X1, X2)
assert cp.array_equal(y1, y2)

Additional context
Add any other context about the problem here.

It seems rand() is used by raft: https://github.com/rapidsai/raft/blob/bf258f39b670d65c60d8a681c8ec45d0ade5690f/cpp/include/raft/random/detail/permute.cuh#L126

[chyunsu3]

I was able to reproduce the bug.

[chyunsu3]

The reproducer script no longer fails if I add shuffle=False argument to make_regression:

from cuml.datasets import make_regression
import cupy as cp

X1, y1 = make_regression(
    n_samples=100, n_features=5, n_informative=5, random_state=42, shuffle=False
)
X2, y2 = make_regression(
    n_samples=100, n_features=5, n_informative=5, random_state=42, shuffle=False
)

# The following no longer throws exception
assert cp.array_equal(X1, X2)
assert cp.array_equal(y1, y2)

So we can conclude that the nondeterminism is introduced by the shuffling process.

[says1117]

I've opened a PR against RAFT to fix the root cause: raft::random::permute was using the unseeded global rand() for its permutation coefficients, so make_regression's shuffle ignored the seed. Once it lands and cuML picks up the new RAFT, make_regression will be reproducible with shuffle=True.

NVIDIA/raft#3075