perf(script-command): avoid full script reads for headers

[JustYannicc]

Summary

• Replace full script reads during discovery with bounded prefix reads for the first Raycast metadata lines.
• Cache shebang interpreter metadata on discovered script commands and reuse it during execution.
• Add a source-backed benchmark plus regression tests for metadata parsing, shebang execution, no-shebang fallback, and large scripts.

Metrics

Benchmark: node scripts/bench-script-command-discovery.mjs with 40 script command files, each with a 1.0 MiB body.

Scenario	Before	After
Discovery time	98.8ms	16.4ms
Discovery runner bytes read	40.0 MiB	1.3 MiB
Cached execution shebang lookup time	8.7ms	4.4ms
Cached execution shebang lookup runner bytes read	1.0 MiB	0 B

Validation

• node scripts/bench-script-command-discovery.mjs
• pnpm test (29 passed, 1 existing live Electron skip)
• pnpm run build:main
• pnpm run check:i18n (exits 0; reports existing Italian locale gaps for settings.ai.whisper.vocabulary and settings.extensions.appSearchScope)
• Codex LSP diagnostics on src/main: 0 errors

Risks

• Header discovery now reads up to the first 120 lines with a 256 KiB cap. That preserves the existing 120-line metadata window for normal Raycast headers, but a script with extraordinarily long header lines could need its metadata moved closer to the top.

[JustYannicc]

Superseded by the consolidated upstream PR #535: #535