Support TIFF predictor=3 on the CPU write path (#1313)

[brendancol]

Summary

• to_geotiff, write_geotiff_gpu, the VRT writer, and the dask streaming writer now accept predictor as bool or int. Passing 3 selects the floating-point predictor (TIFF Technical Note 3), which typically gives better deflate/zstd ratios on float rasters than horizontal differencing.
• Wires fp_predictor_encode (already in _compression.py) into _writer.py's strip, tile, streaming, and _compress_block paths. Adds a normalize_predictor helper that rejects predictor=3 on integer dtypes and preserves the legacy bool semantics: False/True still mean "no predictor" / "horizontal differencing".
• Threads the int through to the GPU writer, which already had a _fp_predictor_encode_kernel waiting to be reached.

Closes #1313.

Test plan

• New tests in xrspatial/geotiff/tests/test_predictor_fp_write_1313.py: round-trip float32 and float64 under deflate and zstd (tiled and stripped), multi-band float32, dask streaming, integer-dtype rejection, legacy predictor=True/False unchanged, compression='none' suppresses the predictor, and a smooth-float file with predictor=3 is smaller than the same file with predictor=2.
• Full xrspatial/geotiff/tests/ suite: 469 passed, 4 skipped, no regressions. (Three pre-existing matplotlib TestPalette failures exist on main and were deselected.)
• On a 512×512 smooth float32 raster, deflate output: predictor=3 590 KB vs predictor=2 787 KB (~25% smaller), confirming the encoder is wired correctly.
• _read_predictor_tag test helper reads the TIFF Predictor tag (id 317) directly from each written file to verify the on-disk tag matches what was requested.