FIX SAGA solver ignoring sample_weight when drawing samples (gh-21305) - #34799

…earngh-21305)

SAG/SAGA draw one sample per inner iteration via
SequentialDataset._get_random_index(), which is uniform over the rows and
never accounts for sample_weight -- sample_weight only scaled the loss of
whichever sample was drawn, not the probability of drawing it. When
weights are far from uniform this makes the stochastic gradient estimate
very high-variance and the solver converges poorly or not at all.

Since sample_weight is fixed for the duration of a solver run, build a
Walker/Vose alias table once (O(n_samples)) so each iteration can draw a
sample proportional to its weight in O(1), and apply the matching
inverse-probability correction to SAGA's direct gradient term to keep the
estimator unbiased (the sum_gradient/intercept_sum_gradient accumulators
are deterministic bookkeeping and must keep accumulating the raw,
unscaled delta).

Restricted to SAGA (is_saga=True): plain SAG has no per-draw bias
correction and relies on every sample being revisited regularly, so
skewing the draw towards heavy samples can slow it down instead of
helping -- a separate, pre-existing limitation this does not address.
Highly skewed weights combined with L1's proximal operator can also still
need more than the default max_iter to satisfy the strict tol criterion,
though the reached solution is markedly better than before at any given
iteration budget.

No-op for uniform sample_weight (including the default, None): the
original code path is taken unchanged, verified bit-identical for a
fixed random_state.