fix(pdf): bound cross-reference parsing

[McanKul]

What

• follow the startxref, /Prev, and /XRefStm chain with a hard hop limit
• parse trailer/xref dictionaries iteratively with token, nesting, name, and string limits
• cap decoded xref streams and validate /W, /Index, /Size, and PNG predictor row allocations before lopdf runs
• match lopdf last-key behavior and ignore fake trailer text inside comments

Why

Malformed xref streams can otherwise trigger large allocations before the read-only text classifier reaches its existing object and content-stream limits.

Verification

• cargo test --manifest-path src-tauri/Cargo.toml --lib (310 passed)
• cargo test --manifest-path src-tauri/Cargo.toml --lib source_content (55 passed)
• 10 focused xref tests, including classic tables, Flate streams, duplicate fields, chained sections, overflow, and predictor limits
• cargo clippy --lib --tests completed; only existing repository warnings remain
• rustfmt --check and git diff --check

Scope

Read-only classifier only: no command, UI, mutation, release, or main changes. Bounded content-operator parsing remains a separate follow-up PR.

Refs #33