feat: rework mine-tasks skill for external users (quick eval mode)

Rewrites the mine-tasks skill from an internal benchmark development
tool to a user-facing "point at your repo, get a baseline-vs-MCP
comparison" workflow. Key changes:

- Add Phase 0 (eval goals): quick eval vs full mining, MCP provider
  selection (Sourcegraph, GitHub Copilot, custom, or placeholder)
- Quick eval mode: mines 5-10 tasks, auto-generates ground truth from
  PR patches, produces standalone run_eval.sh with no Harbor/Daytona
  dependency
- Auto-generate ground_truth.json from PR patch data (files changed,
  patch stats, source PR metadata)
- Generate dual Dockerfiles per task: Dockerfile.baseline (full code,
  no MCP) and Dockerfile.mcp (truncated source + MCP tools)
- Private repo support via docker build --secret for git clone auth
- Language-specific test commands in generated test.sh
- Standalone run_eval.sh runner: builds images, runs agents, collects
  scores, prints baseline-vs-MCP comparison table
- MCP provider is configurable, not hardcoded to Sourcegraph
- Simplified task naming (no internal csb_sdlc_* taxonomy in quick mode)
- Full mining mode preserved for CodeScaleBench contributors

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: sjarmak