Honor per-raster nodata sentinels in merge()

[brendancol]

Closes #1448.

Summary

merge() only detected the first raster's nodata sentinel and reused it for every other input, so rasters with a different sentinel had their invalid pixels passed through reprojection as real data. The mosaic looked fine but contained wrong numerical values.

Fix

• Detect each raster's own input nodata sentinel in the raster_infos loop, independent of the user-supplied output nodata.
• Reproject each raster using its own input sentinel, then canonicalize to NaN before the strategy merge so all rasters share one missing-data marker.
• After the strategy merge, convert NaN back to the user-requested output nodata if that is not NaN.

Applied in both the in-memory path (_merge_inmemory) and the dask path (_merge_block_adapter / _merge_dask).

Test plan

• pytest xrspatial/tests/test_reproject.py (139 passed)
• New TestMergeMixedNodata class covers:
  • NaN + -9999 sentinel mix, default output nodata
  • NaN + 0 sentinel mix with mean strategy (verifies zeros are not folded into averages)
  • Mixed sentinels with explicit user-requested nodata=-9999 output

[brendancol]

@copilot resolve the merge conflicts in this pull request

[Copilot]

@copilot resolve the merge conflicts in this pull request

Resolved in 003a7dd. Combined the per-raster raster_nodata_list from this branch with the same_crs_list optimization that landed on main — both now flow through _merge_block_adapter and _merge_dask. 142 tests pass.