Fix multiarch Docker build and optimize CI

[gricha]

Summary

• Fix Tectonic downloads for both architectures: Use musl-based binaries (statically linked) instead of glibc ones to avoid missing library issues
• Fix Tectonic download URL: Use the versioned release path with proper URL encoding
• Add native arm64 runner: Build arm64 natively on ubuntu-24.04-arm instead of QEMU emulation
• Parallel builds with manifest merge: Build both architectures in parallel on native runners, then merge into multi-arch manifest
• Add BuildKit optimizations: Syntax directive, apt cache mounts, --no-install-recommends

Performance Comparison

Build	Before	After	Improvement
arm64	~38 min (QEMU)	3m49s (native)	~10x faster
amd64	~3 min	5m4s (full rebuild)	Similar
Total	~40 min	~5 min	~8x faster

Root Cause

1. Original arm64 failure: Tectonic only provides aarch64-unknown-linux-musl binaries, not aarch64-unknown-linux-gnu
2. amd64 failure after optimization: --no-install-recommends removed libgraphite2.so.3 needed by glibc binaries

Solution

• Use musl binaries for both architectures (statically linked, no runtime deps)
• Use registry-based caching for faster incremental builds
• Use native arm64 runners for massive speedup

Test plan

• Triggered workflow_dispatch on fix branch
• arm64 build succeeded (3m49s)
• amd64 build succeeded (5m4s)
• Manifest merge succeeded
• Image pushed to ghcr.io

🤖 Generated with Claude Code