Streaming write of monolithic TIFF from dask input

[brendancol]

Author of Proposal: @brendan

Reason or Problem

to_geotiff calls .compute() on dask-backed DataArrays, materializing the
full raster into RAM before writing. For rasters that don't fit in memory, this
makes to_geotiff unusable unless you go through the VRT tiled output path
(see #1083).

Proposal

Add a streaming write path for monolithic .tif output from dask inputs.
Write tiles sequentially as each dask chunk computes, then seek back to patch
the IFD offset/bytecount arrays. This is how GDAL handles it internally.

This is a follow-up to the VRT output work in #1083. The VRT path solves the
immediate problem for most users, but some workflows genuinely need a single
file.

Design: Not yet specced. Needs investigation into how the current
_writer.py assembles IFDs and whether the two-pass approach (write tiles,
then patch offsets) can be grafted onto the existing code without a rewrite.

Drawbacks

• Adds complexity to the writer for a use case the VRT path already covers.
• TIFF offset patching is fiddly and easy to get wrong.

Alternatives

• VRT tiled output (Add dtype, compression_level, and VRT output to geotiff I/O #1083) covers most large-raster use cases.
• Users can write to VRT and convert with gdal_translate if they need a
  single file.