Drop lz4 from byte-parity sweep, gated as experimental (#2268)

[brendancol]

Summary

• xrspatial/geotiff/tests/test_lowlevel_write_pushdown_2138.py::test_write_vs_to_geotiff_byte_parity_uint8[lz4] was failing because _write / to_geotiff now reject lz4 at the experimental-codec gate from geotiff: tier the public surface; mark experimental codecs/features opt-in #2137 unless the caller passes allow_experimental_codecs=True.
• Remove lz4 from _PARITY_CODECS. The other experimental codecs (lerc, jpeg2000, j2k) were never in the tuple, so the result matches the docstring's "stable lossless codec set" framing.
• Note the experimental-codec exclusion in the tuple's leading comment so the next maintainer to add a codec sees the gate up front.

Closes #2268.

Test plan

• pytest xrspatial/geotiff/tests/test_lowlevel_write_pushdown_2138.py -v -- 26 passed (was 26 passed, 1 failed before the change)

[brendancol]

PR Review: Drop lz4 from byte-parity sweep, gated as experimental (#2268)

Blockers (must fix before merge)

• (none)

Suggestions (should fix, not blocking)

• (none)

Nits (optional improvements)

• xrspatial/geotiff/tests/test_lowlevel_write_pushdown_2138.py:46-51: the _codec_available("lz4") branch is now dead code in this file. Nothing else in the module exercises lz4 after the parity-tuple change. You can drop the lz4 branch from _codec_available or leave it as a generic skip-probe for future experimental-opt-in tests. Either is defensible.

What looks good

• Diff is exactly what the issue calls for: one tuple entry removed, with a comment explaining the experimental-codec gate cited to issue #2137.
• The exclusion matches the existing treatment of lerc, jpeg2000, and j2k, which were never in the parity tuple. The list is now internally consistent.
• The remaining 26 parametrizations are stable codecs that round-trip losslessly, so the parity contract the test was written to defend is preserved.
• No production code touched; no test downgraded.

Checklist

• Algorithm matches reference/paper -- n/a, test-only change
• All implemented backends produce consistent results -- n/a
• NaN handling is correct -- n/a
• Edge cases are covered by tests -- existing coverage preserved
• Dask chunk boundaries handled correctly -- n/a
• No premature materialization or unnecessary copies -- n/a
• Benchmark exists or is not needed -- not applicable
• README feature matrix updated (if applicable) -- not applicable
• Docstrings present and accurate -- comment updated to cite #2137 gate