Reject predictor with jpeg2000/lerc/jpeg compression

[brendancol]

Closes #3371

The GeoTIFF writer applied the TIFF horizontal / floating-point differencing predictor (PREDICTOR=2/3) to pixel bytes before handing them to the jpeg2000 and lerc codecs, then stamped PREDICTOR=2 in the IFD. GDAL only honors the PREDICTOR tag for byte-oriented entropy codecs (deflate, lzw, zstd, lz4, packbits), so the result was a file whose pixels read back corrupted in any spec-compliant reader. GDAL read lerc+predictor=2 with a maxdiff of 64537; jpeg2000+predictor=2 would not open at all.

• normalize_predictor already received the compression tag but never used it. It now rejects predictor 2 or 3 combined with jpeg/jpeg2000/lerc with a clear ValueError. Every writer leg (eager strip + tiled, GPU) routes through this function, so the gate covers them all.
• Byte-oriented codecs keep honouring the predictor; the gate is codec-scoped, not a blanket ban.

Backend coverage: eager (numpy) and GPU writer (_write_geotiff_gpu). dask writes route through the same eager encode path.

Test plan

• normalize_predictor raises for predictor 2/3 + jpeg/jpeg2000/lerc; passes through for byte codecs
• to_geotiff raises for predictor=2 + {jpeg2000, j2k, lerc} and predictor=3 + {jpeg2000, lerc}
• to_geotiff predictor=2 + {deflate, lzw, zstd} still round-trips (regression guard)
• GPU writer raises for predictor=2 + {jpeg2000, lerc} (CUDA-gated)
• Full geotiff write suite green (1175 passed)

Note on jpeg

jpeg + predictor=2 previously hit the jpeg branch in _prepare_strip (_encode.py:331), which returns before the predictor branch, so that write silently succeeded with no predictor applied. It now raises, consistent with the jpeg2000/lerc rejection and with GDAL. This is a move from silently-ignore to hard-error for that one combination.

[brendancol]

PR Review: Reject predictor with jpeg2000/lerc/jpeg compression

Blockers

None.

Suggestions

One behavior change worth a line in the PR description (not a code change): jpeg + predictor=2 used to hit the jpeg branch in _prepare_strip (_encode.py:331), which returns before the predictor branch, so that write silently succeeded and produced a file with no predictor applied. It now raises. Raising is the correct, GDAL-consistent behavior, but it moves jpeg+predictor from silently-ignore to hard-error, so it is worth noting so it isn't a surprise.

Nits

• _PREDICTOR_INCOMPATIBLE_COMPRESSION maps each tag to a lowercase name used only in the error message. Fine as is; a set plus a separate name lookup would be more code for no gain. No change needed.

What looks good

• Single chokepoint. All three writer call sites (_writer.py:477, _writer.py:911, _writers/gpu.py:696) pass comp_tag from _compression_tag, so the guard in normalize_predictor covers eager strip, eager tiled, and the GPU writer at once. Confirmed by grep.
• The gate keys on the canonical integer compression tags, which are the only values those call sites ever pass, so it cannot misfire on an unrecognized compression.
• Codec-scoped, not a blanket ban: deflate/lzw/zstd/lz4/packbits + predictor still pass and round-trip (test_to_geotiff_predictor2_byte_codec_round_trips).
• Default writes use deflate, which is byte-oriented, so the common path is unaffected.
• Tests cover the unit level (normalize_predictor), the eager writer end-to-end for predictor 2 and 3, the positive round-trip regression guard, and the GPU writer (CUDA-gated). The predictor=3-requires-float check is preserved.
• The error message names the offending codec, lists the compatible codecs, and tells the caller how to proceed.

Checklist

• Behavior matches GDAL (PREDICTOR only honored for byte-oriented codecs)
• All writer backends covered by one guard (eager strip/tiled + GPU)
• No NaN/dask/compute path touched (validation-only change)
• Edge cases covered: predictor 1/False pass-through, predictor 3 float check kept
• No premature materialization or copies
• Benchmark not needed (validation guard)
• README matrix not applicable (no new function or backend change)
• Docstring updated to note the new rejection