Fix memory leak in liblinear - #34256

There is a memory leak in LinearSVR, as demonstrated by this snippet (run on sklearn v1.9.0, but the issue predates this):

import os
import psutil
import numpy as np
from sklearn.svm import LinearSVR

def get_memory_usage():
    process = psutil.Process(os.getpid())
    return process.memory_info().rss / (1024 * 1024)

X = np.random.randn(1000, 10)
y = X.sum(axis=1)

for i in range(10001):
    if i % 2000 == 0:
        print(f"Iteration {i:5d}: Memory = {get_memory_usage():6.2f} MB")
    LinearSVR(max_iter=10).fit(X, y)

Output:

Iteration     0: Memory = 244.92 MB
Iteration  2000: Memory = 291.26 MB
Iteration  4000: Memory = 337.13 MB
Iteration  6000: Memory = 383.00 MB
Iteration  8000: Memory = 428.87 MB
Iteration 10000: Memory = 474.74 MB

Running under MSAN identified that the problem comes from memory allocated here:

via this line:

but the memory is only freed in the else branch later in the function:

The fix is straightforward: move these free lines outside the else branch, so that they're executed unconditionally.

AI usage disclosure

I used AI to research and understand the issue, but wrote the code in this PR manually.