fix(config): let an explicitly falsy value on the right-hand side win in merge_config

[Rodrigo-Palma]

Rationale

merge_config() says it takes "the non-null value, with precedence on rhs", but the implementation is:

new_config[rhs_key] = rhs_value or lhs_value

or falls back to the left-hand side for any falsy value, not just None. The right-hand side is the caller's own properties, so explicitly turning an option off is silently discarded when the configuration file sets it.

# ~/.pyiceberg.yaml
# catalog:
#   prod:
#     type: rest
#     uri: https://example.com
#     s3.path-style-access: "true"

from pyiceberg.catalog import load_catalog

catalog = load_catalog("prod", **{"s3.path-style-access": False})
# s3.path-style-access is still "true"

The same happens for "" and 0:

rhs value	documented result	actual result
False	False	"true" (from the file)
""	""	"true"
0	0	"true"
None	value from the file	value from the file (correct)

This contradicts what #45 set out to do. That PR swapped the merge order precisely so that "the passed in argument takes precedence" over configuration coming from the environment, and a caller passing False is passing an argument.

There is also a precedent inside this codebase: property_as_bool() in pyiceberg/utils/properties.py already distinguishes "not set" from "falsy", with if (value := properties.get(property_name)) not in (None, ""). After this change the two agree instead of disagreeing.

Change

new_config[rhs_key] = rhs_value if rhs_value is not None else lhs_value

None still means "not set" and still lets the left-hand value survive, which is what the _from_environment_variables path relies on.

Tests added to tests/utils/test_config.py:

• test_merge_config_rhs_wins_for_falsy_values, parametrized over False, "" and 0.
• test_merge_config_lhs_wins_when_rhs_is_none, pinning the None behavior so a future change cannot quietly turn None into an override.

Verification

• make lint passes (ruff, ruff format, mypy, pydocstyle, codespell).
• pytest tests/utils tests/catalog tests/cli passes: 1434 tests.
• pytest tests/ -m "not integration and not s3 and not adls and not gcs" --ignore=tests/io --ignore=tests/benchmark: 3900 passed, 9 failed. The 9 failures are all in tests/avro/test_decoder.py for CythonBinaryDecoder and are pre-existing in my environment: the same 9 fail on the unmodified branch at the same commit (9 failed, 59 passed both with and without this patch), since the Cython extension is not built locally.