Optimize svgbob rendering hot paths

[exlee]

I am evaluating svgbob as a library backend for my DiagramIDE app.
Before integrating it as another diagram output language, I profiled the
renderer (current Pikchr output is mindboggingly fast) and found some scaling
problems on larger connected diagrams.

I used GPT-5.6 as an optimization partner while investigating the hot paths, and
wanted to share the concrete results with the upstream project.

I do not have any expectations toward handling this PR. If it's not aligning
with project direction or personal views, feel free to disregard it - I merely
found a value in it and thought it's share-worthy.

Results

Release build benchmark using svgbob_cli, with three warmup runs and eight
timed runs per fixture:

Fixture	Before	After	Improvement
simple.bob	25.7 ms	9.1 ms	2.8x
demo.bob	864.1 ms	50.9 ms	17.0x
long.bob	2037 ms	152.9 ms	13.3x
circles.bob	167.9 ms	17.7 ms	9.5x
example.bob	295.9 ms	39.3 ms	7.5x

The generated SVGs were byte-for-byte identical for all five fixtures.

What changed

• Replaced recursive all-pairs span merging with indexed union-find
  connectivity over neighboring cells, while preserving input order.
• Cached fragment bounds when constructing FragmentTree nodes instead of
  recalculating geometry bounds during every containment comparison.
• Localized circle and arc candidates once per search span.
• Made failed circle/arc subset checks return immediately when a required
  template cell is absent, avoiding the second scan for unmatched cells.
• Added a regression test covering transitive span connectivity.
• Refreshed the stale lockfile package metadata from 0.7.5 to 0.7.6.

Verification

• cargo test --offline -p svgbob
  • 105 unit tests passed
  • 5 simple_shapes integration tests passed
  • styling integration test passed
  • doc tests passed or remained intentionally ignored
• cargo fmt --all -- --check passed.
• Release benchmarks and SVG byte comparisons passed.

cargo clippy --offline -p svgbob --all-targets -- -D warnings still reports
24 existing upstream lint findings in untouched code; none are introduced by
these changes.

LLM Disclaimer

I've set search boundary condition, harness, search requirements and let GPT 5.6 Sol/Luna run.
There wasn't much work on my side toward this contribution, outside of preparing this PR and validating result.