fix: prevent benchmark alignment look-ahead

[JohnDReynolds]

Summary

Prevent benchmark alignment from moving a future benchmark return onto an earlier strategy date.

When a strategy calendar contains dates missing from the benchmark, _prepare_benchmark() reconstructs benchmark prices on a daily index before converting them back to returns. Back-filling that price index allowed a later observed benchmark price to propagate backward, creating look-ahead in reports, plots, and statistics that consume the aligned benchmark.

This changes the alignment from backward-fill to forward-fill so each benchmark return remains on its observation date.

Behavior

Consider a benchmark observed on Friday and Monday, with a 10% return on Monday, aligned to a daily strategy calendar:

Dates:     Friday  Saturday  Sunday  Monday
Previous:   0.00      0.10     0.00    0.00
Expected:   0.00      0.00     0.00    0.10

Back-filling assigned Monday's price to earlier weekend dates, causing the return to appear before it occurred. Forward-filling carries Friday's last known price through the weekend and preserves the return on Monday.

The change preserves:

• already aligned benchmark data;
• leading and trailing strategy dates;
• positive and negative benchmark returns;
• compounding across irregular target periods;
• Series and one-column DataFrame inputs;
• labels, indexes, and timezone normalization; and
• caller-owned benchmark data.

Regression evidence

• Fail before: A Friday/Monday benchmark aligned to a daily strategy calendar produced [0, 0.10, 0, 0] instead of [0, 0, 0, 0.10]. Through the public statistics path, information_ratio() returned -0.261116483933547 instead of the independently calculated -0.3100868364730212.
• Independent expectation: Alignment must be causal: a benchmark price may be carried forward from its last observation, but a later price must never be propagated to an earlier date.
• Pass after: The return remains on Monday for either sign; leading and trailing dates remain flat; intervening irregular-period returns compound to 0.32; timezone-aware DataFrame alignment preserves labels and caller state; and information_ratio() matches its independent active-return calculation.

Verification

• 5 focused causal-alignment tests passed.
• 152 utility, statistics, and regression tests passed.
• 35 downstream report and plot tests passed.
• 230 full-suite tests passed with 60 inherited seaborn/Matplotlib deprecation warnings.
• Changed-line and full Ruff lint passed.
• Ruff formatting and git diff --check passed.
• No public signature, documentation, dependency, packaging, or build behavior changed.

Scope

Changed files:

• quantstats/utils.py
• tests/test_regressions.py

The production change is limited to replacing backward price filling with forward price filling inside _prepare_benchmark().

Out of scope:

• changing benchmark statistic formulas;
• redesigning end-of-year reports;
• changing benchmark downloads; and
• unrelated alignment or formatting cleanup.