Restructure benchmark results table to reduce row count

[BryanFRD]

Problem

The current benchmark table is a flat list with one row per fixture/tool/method/command combination. With ferrflow alone that's already 48 rows (4 fixtures × 3 methods × 4 commands). Adding changesets, semantic-release, and release-please will push this to 192+ rows — unreadable in a PR comment or release note.

Current format:

Fixture	Tool	Method	Command	Median	Memory
mono-large	ferrflow	binary	check	189.8ms	21.5 MB
mono-large	ferrflow	binary	release-dry-run	187.5ms	21.4 MB
mono-large	ferrflow	binary	tag	1781.8ms	15.1 MB
mono-large	ferrflow	binary	version	12.3ms	10.9 MB
...	...	...	...	...	...

Proposal

Group by method (binary / npm / docker), then use a pivot layout where commands become columns:

Binary

Fixture	Tool	check	release	tag	version	Peak RSS
single	ferrflow	14.7ms	13.7ms	13.0ms	10.8ms	12.4 MB
mono-small	ferrflow	15.1ms	15.0ms	15.6ms	11.2ms	12.6 MB
mono-medium	ferrflow	23.2ms	23.6ms	70.5ms	11.6ms	13.0 MB
mono-large	ferrflow	189.8ms	187.5ms	1781.8ms	12.3ms	21.5 MB

This reduces 48 rows → 12 rows for one tool, and 192 rows → 48 for four tools. Each table stays compact enough for PR comments.

When multiple tools are present, they sort together per fixture for direct comparison:

Fixture	Tool	check	release	tag	version	Peak RSS
mono-large	ferrflow	189.8ms	187.5ms	1781.8ms	12.3ms	21.5 MB
mono-large	semantic-release	...	...	...	...	...
mono-large	changesets	...	...	...	...	...

Scope

• scripts/run.sh — main table generation
• scripts/format-release.sh — release notes summary
• Possibly scripts/compare.sh if regression output uses the same format