test: cover file ingestion and round-trip safety for clean_csv/clean_excel (#478)

Adds tests/test_ingestion_roundtrip.py, pinning what survives the reader
boundary in fd.clean_csv and fd.clean_excel — the dtypes a caller gets back
are decided by pandas before any cleaning step runs.

Covered:
  * Excel serial dates (45000) are read as plain integers; only a
    date-formatted cell becomes a timestamp, and the column name is no hint.
  * duplicate headers in CSV and Excel (pandas "name.1" -> "name_1"), plus
    labels that collide only after normalization ("Name"/"name "/"NAME"),
    and the stability of those names across a save and reload.
  * cp1252/latin-1 input: the bare UnicodeDecodeError with no encoding
    option, accented text preserved when an encoding is passed, C1 control
    characters from a latin-1 misread, and mojibake that survives cleaning
    untouched while fd.lint_text_encoding flags it.
  * raw -> clean -> save -> reload -> clean for CSV and Excel, asserting the
    second clean is a no-op down to the written bytes, including quoting,
    embedded newlines and empty cells.
  * formula sanitizing: the returned frame is never rewritten, the written
    file is, so cycle 1 -> 2 changes values once and is idempotent after;
    sanitize_formulas=False round-trips byte-exactly.
  * malformed and degenerate files: ragged long/short rows, header-only and
    empty inputs.

Two assertions pin behaviour that loses information and say so inline:
preserve_leading_zeros is a no-op for clean_excel (pandas turns the text
cell "02134" into 2134, so a CSV -> Excel hop undoes what clean_csv kept),
and the leading-zero pre-scan stops at LEADING_ZERO_SCAN_ROWS.

Author: kevincostner17