Add opt-in winsorization for revenue/price outliers

[BasLinders]

Summary

• New prox.winsorize_series() in prox/data_manager.py: same technique as first-order-engine's ContinuousMetricEngine.winsorize_series (https://github.com/BasLinders/first-order-engine/blob/main/foe/continuous/operations.py#L110) — caps values at mean ± N standard deviations, or at a percentile band — adapted to return a pd.Series instead of a JSON-serializable list, to match this codebase's DataFrame-in/DataFrame-out convention (PRoX doesn't depend on foe outside the optional BigQuery extra, so this isn't imported from there).
• Wired into main.py as a new opt-in "2. Handle Outliers" step (existing sections renumbered 3→6 to make room), applied to the revenue/price column on both raw_df and df_ready right after the data is loaded/cached and before any filtering, sampling, or analysis reads that column — caps values in place rather than dropping the rows they came from.
• Off by default, with a method choice (Standard Deviation / Percentile) and a slider for the threshold, plus a summary of how many values were capped and to what range.

Why

Average Order Value, revenue trend, and category revenue breakdown in Business Insights all read from the same price/revenue column, so a handful of extreme values (e.g. one bulk order) can dilute those stats with no way to address it short of editing the source data. Scoped to the revenue/price column specifically — winsorizing case duration/lead time directly would mean mutating event timestamps, which is a separate, more invasive change.

Test plan

• pytest tests/ — 135 passed (6 new tests: percentile capping of a single injected outlier, std-based bounds match mean ± N·std, NaN positions preserved, no-outlier case leaves values unchanged, empty/all-NaN series return zero bounds)
• Manually injected an extreme outlier (999999) into generated mock data and confirmed both methods cap it correctly (percentile bounds stayed tight and sane; std bounds showed the technique's known sensitivity to a single huge outlier inflating its own reference std — same behavior as the upstream first-order-engine implementation, not a regression)
• Ran the full pipeline end-to-end with winsorized price fed into run_full_analysis and confirmed Average Order Value reflects the capped values
• Manual click-through of the new checkbox/method/slider controls in a running Streamlit session — not run in this environment (no display)