Add error bars to line and scatter plots (`.error_bar()`)

[MoonFlowww]

Summary

Add error bar rendering to line plots and scatter plots — matching matplotlib's plt.errorbar(). This is essential for any benchmark visualization that needs to show variance, standard deviation, or confidence intervals.

Motivation

When comparing algorithm implementations, the mean alone is misleading. Error bars communicate the spread (e.g., ±1 stddev, min/max, or 95% CI) directly on the plot. Without this, Sepia cannot produce publication-quality performance charts.

Proposed API

std::vector<double> x      = {1, 2, 3, 4};
std::vector<double> y      = {10.2, 15.7, 9.8, 13.1};
std::vector<double> y_err  = {0.5, 1.2, 0.3, 0.8};   // symmetric ±
// Or asymmetric:
std::vector<double> y_lo   = {0.4, 0.9, 0.2, 0.7};
std::vector<double> y_hi   = {0.6, 1.5, 0.4, 0.9};

figure.plot(x.data(), y.data(), 4)
      .error_bar(y_err.data())                         // symmetric
      // or:
      .error_bar(y_lo.data(), y_hi.data())             // asymmetric
      .color(sepia::Color{60, 140, 255})
      .label("mean latency");

Implementation Notes

• Store optional error arrays inside PlotEntry (nullable pointers + ownership flag)
• During render_data() in figure.hpp, after drawing the main line/markers, iterate points and draw vertical cap-lines: a vertical segment of height 2*err centered on the point, plus horizontal caps of fixed pixel width (e.g., 4px)
• X error bars follow the same pattern horizontally
• Cap width should be a style parameter (default: 4px)
• CoordTransform already handles the data→pixel mapping needed

Acceptance Criteria

• Symmetric Y error bars render correctly
• Asymmetric Y error bars render correctly
• X error bars supported
• Cap size configurable
• Works on both plot() and scatter() series
• Example added to examples/error_bars.cpp