feat: improve skill_fitness_metric with multi-dimensional scoring

[vominh1919]

Summary

Replaces the single keyword-overlap scorer in skill_fitness_metric() with a weighted composite of five independent signals that spread scores across a much wider range.

Fixes #12

Problem

The original metric used only keyword overlap (len(expected & output) / len(expected)), producing scores in a narrow 37-49% band regardless of actual output quality.

Solution

Five scoring components, each providing independent signal:

Component	Weight	Purpose
Keyword overlap	25%	Stop-word filtered, F1-style recall/precision blend
Char 3-gram similarity	25%	Jaccard on character shingles — captures partial/substring matches
Structural pattern match	20%	Checks for code blocks, lists, headers, bold, URLs
Length quality	15%	Penalizes outputs too short or too long vs. expected
Content density	15%	Unique token ratio, avg token length, sentence variety

Additional changes

• Returns dspy.Prediction(score=float, feedback=str) for GEPA reflective mutation compatibility
• Feedback string highlights specific weaknesses for the optimizer
• All scoring is deterministic (no LLM calls) — fast for optimization loops

Testing

The new metric produces varied scores across different output qualities rather than clustering in a narrow band.