MD051: Replace regex parsing with native tree-sitter link analysis

[ekropotin]

Problem

The current MD051 implementation uses regex to parse links within tree-sitter nodes, which is redundant and architecturally suboptimal.

Current flawed approach:

1. Tree-sitter parses the document into structured nodes
2. We extract text from inline nodes
3. We re-parse the text with regex: r"\[([^\]]*)\]\(([^)]*#[^)]*)\)"
4. We manually calculate positions using custom byte_to_point conversion

Tree-Sitter Already Provides Link Structure

Tree-sitter markdown parser already gives us precise link structure:

paragraph: "[test link](#fragment)" 
  inline: "[test link](#fragment)" (0:0-0:22)
    [: "[" (0:0-0:1)           // Link start
    ]: "]" (0:10-0:11)         // Link text end  
    (: "(" (0:11-0:12)         // URL start
    #: "#" (0:12-0:13)         // Fragment marker
    ): ")" (0:21-0:22)         // Link end

Proposed Solution

Replace regex parsing with native tree-sitter traversal:

1. Find link patterns in tree-sitter nodes: sequence of [ → text → ] → ( → URL → )
2. Extract fragments from URL tokens that contain #
3. Use tree-sitter positions directly (no byte_to_point conversion needed)
4. Filter external links by checking if URL starts with protocol/path before #

Benefits

• Eliminate regex dependency for link parsing
• Remove custom position calculation (use tree-sitter's precise positions)
• Better edge case handling (tree-sitter is more robust than regex)
• Performance improvement (no redundant parsing)
• Cleaner architecture (single source of truth for document structure)

Implementation Notes

The tree-sitter approach would be more aligned with how other rules work in the codebase and eliminate the architectural inconsistency of mixing tree-sitter and regex parsing.

Current regex-based approach works correctly but is architecturally suboptimal.