Improve token estimation for comparison prompts

[ArjenSchwarz]

Context

From code review of #21 (multi-spec comparison feature).

Current Behavior

The estimatePromptTokens function in internal/comparison/prompt.go:87-89 uses a simple heuristic of 4 characters per token:

func estimatePromptTokens(prompt string) int {
    return len(prompt) / 4
}

Issue

This is a rough estimate. Claude's tokenizer typically uses ~3.5-4 chars/token for English prose, but code diffs may have different characteristics (more symbols, varied line lengths).

Suggested Improvement

Options:

1. Use a more conservative estimate (e.g., len(prompt) / 3) for safety margin near the 150k token limit
2. Use a proper tokenizer library for accurate estimates
3. Add different heuristics for code vs prose content

Priority

Low - current implementation works but could fail on edge cases near the context limit.