fix(scripts): protect latency CSV from cache path collisions

[skyc1e]

Passing the same file to --csv and --db with --rebuild-db deletes the source before reading it. With a valid header-only CSV, the command even exits successfully after replacing the input with a SQLite database.

Reject matching file identities before opening the cache or rebuilding it, including relative paths, symlinks, and hardlinks. The CLI now reports --csv and --db must refer to different files, exits with status 1, and leaves the source unchanged. Normal reporting with a separate cache is preserved.

Validation:

• cd scripts && python -m unittest test_tx_latency_report -q: 54 tests, successful with one symlink subtest skipped because Windows requires additional privileges. The hardlink case passes.
• Added seven regression/success tests; the rejection cases fail against the original script.
• Confirmed the real CLI changes from exit 0 with overwritten input to exit 1 with identical source bytes, using a temporary CSV.
• git diff --check passes. Tested with Python 3.12.14; Rust/node integration tests were not run for this Python-only change.