MRB-648 Scorecards for evalml

[adestefani8]

What this PR adds

This PR adds a new report_scorecard rule that renders a PNG comparing one run against one baseline.

The scorecard has:

• one row per (variable × metric)
• one column block per region
• one column per lead time inside each region block

Each cell encodes the row's metric as a relative difference between the two runs for that region and lead time: (model − baseline) / |baseline| × 100.

Markers:

• blue → model better
• red → baseline better
• grey → |diff| below the neutral threshold (default 5%)
• grey x → missing or non-finite value

Scores:

• Supported scores: RMSE, MAE, STDE, R2, ETS, POD, FAR
• Score direction: RMSE, MAE, STDE, and FAR are lower-is-better; R2, ETS, and POD are higher-is-better.

Above the neutral threshold, dot area scales linearly with |diff|% and caps at size_cap_pct (default 30%).

Configuration

Configurable via params on the rule:

• lead_times: "start/stop/step" in hours.
• regions: regions to include as column blocks. If no region is specified, all regions are included.
• variables: "VAR:M1,M2,..." entries. Omit :M1,M2,... to use all_metrics for that variable.
  Metric names can also expand by prefix: requesting ETS includes all matching categorical scores, such as ETS_gt_0p0, ETS_gt_0p001, ETS_gt_0p005.
  If no variable is specified, the script falls back to RMSE only for a default set of variables.

Other defaults (season, init_hour, metric settings, plot styling) live in the script's cfg.

Plot layout

The plotting script makes a few automatic layout decisions:

• the longest region label is measured before rendering: col_width grows when necessary to prevent region header overlap, and the top margin/vertical separators adapt to the rendered header height
• the longest metric label is measured before rendering: variable labels keep a fixed gap from metric labels, and horizontal group separators start from the measured metric-label area
• the legend is centered on the scorecard area
• the no-data legend entry only appears when missing values are present

TODOs

• Expose main scorecard parameters in the evalml config
• merge Add ExperimentConfig for experiment workflow outputs #160

[dnerini]

looking good :)

[jonasbhend]

Very nice. I really like the scorecards. Great work.

For future PRs, could you please add a short description of the changes (high-level overview) and - if necessary - also of the goals of the PR? That would be very helpful for the review.

As an additional suggestion, could we include the scorecard in the dashboard (I know we don't always want to produce it, but in case it is available it would be nice to include in a separate tab)?

[jonasbhend]

i have just left two minor comments. Seems all good to me otherwise.

[dnerini]

let's merge this, thank you all for all the precious feedback and comments!