docs(compress): follow the compress.target rework

[a-essawy]

The compress page described behaviour that changed in rendobar/rendobar#722, which is live in production.

• Video tiers are VMAF 94 / 91 / 86 on the NEG model, and visually_lossless starts at 94.
• The report example gains scoredOn, with a line on what full and sample mean.
• Passthrough reports a null score and level, and a dry run predicts it.
• Byte budgets on H.264 video: one two-pass encode, resolution step-down, probes: 1, qualityLevel: null, AAC audio, untouched when the source already fits.
• JPEG to JPEG XL pixel-exact repack and its mode: "lossless" result.

Frontmatter and nav validators pass locally.