Fix ESMF read casting elementConn before checking for padding

[cmdupuis3]

Closes #1747

Overview

_read_esmf casts elementConn to INT_DTYPE and only then compares against INT_FILL_VALUE, so the padding is recognized in neither of the two ways xarray can hand it over. With CF decoding on (mask_and_scale=True, the default), xarray replaces ESMF's -1 with NaN and promotes the array to float; casting NaN to an integer dtype is platform-dependent, and on arm64 it yields 0, so the padding decodes to -1. With decoding off, the raw -1 survives the cast and decodes to -2. Both are negative indices that silently wrap to the end of the coordinate arrays instead of being treated as fill.

The fix locates the padding before the cast, from numElementConn — the format's own statement of each face's node count — and additionally treats NaN as padding whatever that count claims.

The test builds the ESMF dataset by hand instead of going through _encode_esmf, so it exercises the reader in isolation and covers third-party ESMF files rather than only our own output. It is parametrized over mask_and_scale to pin both decoding paths.

PR Checklist

General

• An issue is created and linked
• Added appropriate labels (if your uxarray repo permissions allow it)
• Filled out Overview and Expected Usage (if applicable) sections

Testing & Benchmarking

• There is adequate test coverage of changes from this PR (add new tests if needed)

Documentation and Examples

• Docstrings updated with any function changes, and included in all new functions
• User (public) functions added to docs/api.rst; internal (private) function names start with an underscore (_)

AI Disclosure

AI Usage: Claude Opus 5

• I have tested and take responsibility for all AI-generated content in my PR.