Fix first-window maximum in Max Subarray Sum reference

[frezcirno]

The Max Subarray Sum reference loses the first window when it is the unique maximum. For example, input=[9, -8], N=2, window_size=1 currently produces -8 instead of 9, so a correct solution can be rejected.

max_sum = current_sum aliases the same scalar Tensor. The first in-place sliding-window update changes both values before torch.max can preserve the original maximum. Clone the initial sum to keep it independent. Add functional regressions for positive, negative, and multi-element first-window maxima, plus the single-window boundary (window_size == N).

Validation:

• Example, 10 functional cases, and the N=50000/window_size=25000 performance case all match an independent Python integer sliding-window oracle on CPU and CUDA (A800, PyTorch 2.5.1+cu121), with three calls per case.
• The unmodified reference fails the three new first-window regressions on both devices; the fixed reference passes.
• python scripts/validate_challenges.py: 101 challenges, zero errors; two existing warnings about unrelated challenge directories/starters.
• Black 24.1.1, isort 5.13.2, flake8 7.0.0 pass for the changed file; git diff --check passes.