Add regression test for GPU pred=3 multi-sample TIFFs #1479

Summary

The GPU _fp_predictor_decode_kernel path in _gpu_decode.py decodes multi-sample float TIFFs (predictor=3, samples=3 or 4, float32 or float64) correctly. Manual round-trip testing confirms it works. But xrspatial/geotiff/tests/test_predictor_multisample.py only covers GPU predictor=2 multi-sample. The predictor=3 multi-sample GPU path is not exercised in CI.

This is a coverage gap, not a bug. The code is right; the suite just doesn't lock it in.

Proposal

Add parametrized regression tests over samples in (3, 4) and dtype in (float32, float64) that:

1. Generate random multi-sample float data of shape (64, 64, samples).
2. Round-trip through to_geotiff(compression='deflate', predictor=3, tiled=True, tile_size=32).
3. Read back on CPU and GPU.
4. Assert GPU is bit-exactly equal to CPU and both match the source.

Use the existing gpu_only skipif decorator.