WithinSession/WithinSubjectSplitter overwrite an explicit n_splits (and shuffle/random_state) in cv_kwargs

[bruAristimunha]

Summary

WithinSessionSplitter.__init__ (and WithinSubjectSplitter) overwrites an explicitly
passed n_splits (and shuffle / random_state) in cv_kwargs with its own n_folds,
unconditionally:

self._cv_kwargs = dict(**cv_kwargs)
params = inspect.signature(self.cv_class).parameters
for p, v in [("n_splits", n_folds), ("shuffle", shuffle), ("random_state", self._rng)]:
    if p in params:
        self._cv_kwargs[p] = v      # clobbers a value the caller already put in cv_kwargs

So a caller can't ask for a different fold count than n_folds even when the inner CV exposes
n_splits.

Repro

from sklearn.model_selection import StratifiedShuffleSplit
from moabb.evaluations.splitters import WithinSessionSplitter

sp = WithinSessionSplitter(cv_class=StratifiedShuffleSplit, n_splits=1, test_size=0.25,
                           shuffle=True, random_state=42)
# wanted: a single 75/25 stratified split per session; got: 5 folds (n_splits overwritten by n_folds)

Use case / current workaround

We want a single stratified train/test split at a paper-reported ratio. Because the explicit
n_splits=1 is clobbered, we wrap StratifiedShuffleSplit in a class that simply doesn't expose
n_splits (so the loop above skips it and the inner split count stays 1). A no-op wrapper that only
exists to dodge the overwrite.

Proposed

Only inject the default when the caller hasn't already provided it:

for p, v in [("n_splits", n_folds), ("shuffle", shuffle), ("random_state", self._rng)]:
    if p in params and p not in cv_kwargs:
        self._cv_kwargs[p] = v

Backward compatible (defaults unchanged when cv_kwargs doesn't set these), and lets a single
holdout split be requested via plain cv_class=StratifiedShuffleSplit, cv_kwargs={n_splits:1, ...}.
Happy to PR.