Refactor GeoTIFF Phase 5f: extract _encode.py from _writer.py

[brendancol]

Closes #2260
Part of #2211

Summary

Move the write-side encode helpers out of _writer.py into a new
_encode.py. Mirrors the _decode.py extraction on the read side
(PR-G).

Helpers relocated:

• Strip and tile encode: _prepare_strip, _write_stripped,
  _prepare_tile, _write_tiled
• Streaming block compression: _compress_block
• Photometric and MinIsWhite: _resolve_photometric,
  _reject_disagreeing_photometric_override,
  _apply_photometric_miniswhite_invert,
  _invert_nodata_for_miniswhite,
  PHOTOMETRIC_MINISBLACK / PHOTOMETRIC_RGB /
  _PHOTOMETRIC_NAME_MAP
• Predictor: normalize_predictor, _apply_predictor_encode
• Compression name mapping: _compression_tag
• Threshold constant: _PARALLEL_MIN_BYTES

_writer.py keeps _write, _write_streaming,
_validate_lowlevel_write_kwargs, and the fsspec write helpers, and
re-exports every moved name so existing import paths and the
_writers subpackage continue to work without churn.

_write_tiled and _write_stripped look up _prepare_tile and
_prepare_strip through _writer so the monkeypatch contract
pinned by test_gil_friendly_kwarg_1830 survives the move.

_writer.py drops from 1800 to 1269 lines (~531 line move).

Test plan

• pytest xrspatial/geotiff/tests/ -- 5037 passed, 68 skipped,
  one pre-existing lz4 byte-parity failure unrelated to this PR.
• Photometric / predictor / strip / tile / miniswhite / compression
  tag tests pass unchanged.
• Public import surface (from xrspatial.geotiff._writer import _resolve_photometric etc.) still resolves.

[brendancol]

PR Review: Refactor GeoTIFF Phase 5f: extract _encode.py from _writer.py

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

• _encode.py line 27-29 (module docstring): the docstring says heavier helpers are looked up lazily inside _compress_block / _prepare_strip / _prepare_tile, but the lazy _writer import is actually in _write_stripped (line 369) and _write_tiled (line 516). _compress_block, _prepare_strip, and _prepare_tile themselves do no lazy lookup. Tighten the wording to point at the actual call sites.

Nits (optional improvements)

• The lazy-lookup comments at _encode.py:365-368 and _encode.py:511-515 are clear, but consider adding a one-line cross-reference to the test that pins the contract (test_gil_friendly_kwarg_1830::test_write_tiled_parallel_passes_gil_friendly_positionally), so the next refactor knows which test will catch a regression.

What looks good

• Move is mechanical and byte-neutral. Every helper and constant lands in the same shape it had in _writer.py.
• Re-export block at _writer.py:97-114 preserves the historical import surface for the _writers subpackage, _gpu_decode, and the test suite.
• Lazy _writer._prepare_tile / _writer._prepare_strip lookup inside _write_tiled and _write_stripped keeps the monkeypatch contract pinned by test_gil_friendly_kwarg_1830 working without test changes. Same pattern _write_layout.py uses for _resolve_photometric.
• No module-level cycle: _encode.py only pulls _compression, _header, and _write_layout at load time; _writer is imported lazily inside the two orchestrators that need the monkeypatch indirection.
• _writer.py line count: 1800 to 1269 (-531). Inside the issue's ~1300 target.
• Module docstring follows _decode.py's style and calls out the symmetry with PR-G.

Checklist

• Byte parity preserved (no algorithmic changes)
• Public import paths preserved (xrspatial.geotiff._writer.<name> resolves for every moved name)
• Monkeypatch contract preserved (test_gil_friendly_kwarg_1830, test_streaming_write_parallel, test_write_layout_monkeypatch_contract_2248 all still pass)
• Full geotiff test suite passes (5037 passed; one lz4 failure is pre-existing on main, not caused by this PR)
• No new public API
• No README / docs touch needed (private refactor)

[brendancol]

PR Review (second pass): Refactor GeoTIFF Phase 5f: extract _encode.py from _writer.py

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

None remaining. Prior pass's docstring suggestion was applied at _encode.py:23-32.

Nits (optional improvements)

None remaining. Prior pass's cross-reference nit was applied at _encode.py:365-370 (strip path) and _encode.py:511-518 (tile path, with the full test node-id).

What looks good

• Follow-up commit is documentation-only; behaviour unchanged.
• Module docstring now names _write_stripped and _write_tiled as the two lazy-lookup sites instead of pointing at the three prepare/compress leaves.
• Both lazy-lookup comments now name the pinning test by id, so a future refactor that inlines the indirection will know which test will catch it.
• Full geotiff test suite still passes the way it did on the first commit (5011 passed excluding the pre-existing lz4 byte-parity failure on main).

Checklist

• Byte parity preserved
• Public import paths preserved
• Monkeypatch contracts preserved (test_gil_friendly_kwarg_1830, test_streaming_write_parallel, test_write_layout_monkeypatch_contract_2248)
• Full geotiff test suite passes (modulo the pre-existing lz4 failure)
• No new public API
• Documentation matches actual call sites

Clean. No further changes requested.