feat(M83): Divergence Heatmap by Token Position

[hlin99]

Problem

When analyzing batch comparison results, it's unclear whether divergences cluster at specific token positions (early prefill vs late decode).

Solution

Add xpyd-acc heatmap --report <path> command that:

• Analyzes divergence frequency by token position across all samples
• Bins token positions into configurable buckets (default 10)
• Per-bucket: divergence count, divergence rate, avg logprob gap
• HeatmapBucket and HeatmapReport dataclasses with to_dict() serialization
• --buckets <int> flag, --json <path> export
• Rich terminal bar chart showing divergence frequency by position range

Acceptance Criteria

• heatmap.py module with compute_heatmap(), HeatmapReport, format_heatmap()
• CLI subcommand heatmap with --report, --buckets, --json flags
• Rich bar chart output
• 12+ tests covering bucket computation, edge cases, formatting, JSON export, CLI
• All lint/tests pass