fix(stats): end box plot whiskers at observed inliers

[Shubham-Padkonde]

Fixes #1731.

When a sample contains outliers, computeStats currently returns the theoretical Tukey fences as box-plot whiskers, even if no observation lies there. For [-100, 1, 2, 3, 4, 5, 6, 7, 8, 9], this produces -5.5 and 14.5 instead of 1 and 9.

Use the first and last sorted non-outlier observations for the box-plot endpoints. Reuse the observations already selected for binning; histogram bounds, bin counts, quartiles and outlier classification retain their existing behavior.

Regression coverage includes lower, upper and two-sided outliers, values exactly on the fences, input preservation and unchanged histogram bins. The three outlier cases fail on the base revision and pass with this change. The zero-IQR work in #2024 remains separate.

Validation (Windows, Node 24.15, Yarn 4.10.3):

• All 15 @visx/stats tests passed.
• Changed-file ESLint and Prettier passed without warnings or errors.
• Workspace TypeScript builds across 44 packages and kernel API type check passed.
• Vendor compilation and CommonJS/ESM package compilation passed. The top-level build's Unix find cleanup and the package helper's inline environment assignment do not run in Windows shells; I ran the remaining type checks and Babel steps directly without changing build scripts.

Prepared with Codex assistance.