Preserve integer precision in mixed numeric DataFrames - #459

Fixes #240.

A DataFrame containing integer and floating columns can expose a shared floating-point .values matrix. Integers are converted before formatting: intfmt is ignored and values above the exact floating-point range can round, for example 2**53 + 1.

When that matrix has a floating dtype and an integer column is present, read plain row tuples with itertuples(index=False, name=None). Preserve existing index/header handling and column order, including duplicate names, and retain the matrix path for other inputs. Tests cover large signed/unsigned values, indices, column labels, empty dimensions, nullable values, mixed text, and temporal controls. The README documents the bounded numeric-column behavior.

Validation rerun on Windows 11 / Python 3.11.9 / NumPy 2.4.6 / pandas 3.0.6:

• python -m pytest -q: 413 passed, 1 skipped, including doctests.
• Ruff 0.15.4 check and format --check: passed.
• A representative added formatting regression fails against the unchanged upstream base.
• git diff --check: passed.

Ruff 0.16.10 was also attempted and reported newer lint/README formatting diagnostics; the passing lint run uses the project's declared development minimum, 0.15.4. No source was changed to resolve unrelated tool-version diagnostics.

AI assistance was used to prepare and verify the patch. Archived Linux version-matrix/build/benchmark checks were not rerun here. Numeric categorical/complex inputs and global scalar inference are outside the fix. It is a correctness change with possible fixed overhead on small DataFrames; no new performance claim is made. Submitted for maintainer review and upstream CI.