Fix local models and render terminal images natively

[hrhrng]

What changed

• fix managed voice-model downloads so Hugging Face snapshots use the correct repository argument, disk estimates use the right units, and downloaded models can be removed with a double-d confirmation
• make /forget remove persisted reading progress, speech checkpoints, and the bare-launch resume target as well as the library entry
• add real import, reading, and local MOSS read-aloud media to the README
• render EPUB covers and illustrations through native Kitty, iTerm2, or Sixel protocols, including images clipped by scrollback
• reject the Unicode half-block fallback so unsupported terminals show an explicit placeholder instead of a pixelated mosaic
• update Ratatui/Crossterm and the documented Rust requirement for the native image stack

Why

The model installer could briefly report downloading and then return to not downloaded because the generated Hugging Face call used the wrong argument. Separately, illustrations were always downsampled into Unicode half blocks even when the terminal supported exact pixel graphics. /forget also left the book's content-keyed state behind, so reimporting the same file resumed the old session and suppressed the fresh-book cover.

Validation

• cargo fmt --check
• cargo test --locked --test image --test illustration (18 passed)
• cargo test --locked --test render (27 passed)
• cargo test --locked --lib (244 passed)
• cargo test --locked --test library (11 passed)
• cargo test --locked (all completed batches passed; one timing-sensitive render test timed out once and passed both isolated and full-suite reruns)
• cargo build --release --locked
• installed the release binary locally and verified its SHA-256 matches the build artifact