Geotiff readers/writers reject file-like (BytesIO) sources despite advertising 'str | path-like'

[brendancol]

Problem

open_geotiff, read_geotiff, read_geotiff_dask, read_geotiff_gpu, read_vrt, and to_geotiff are all typed as taking source: str / path: str, and the docstrings describe a file path. The reader's _open_source (xrspatial/geotiff/_reader.py) only handles local paths, http(s) URLs, and fsspec URIs — passing a BytesIO (or any file-like) raises TypeError/AttributeError.

This is a common ask for in-memory pipelines (test fixtures, fetch-then-decode, S3 multipart staging, network-attached buffers) and the obvious workaround (write to a temp file) defeats the purpose of having a fast streaming reader.

Repro

```python
import io, numpy as np, xarray as xr
from xrspatial.geotiff import to_geotiff, read_geotiff

buf = io.BytesIO()
to_geotiff(xr.DataArray(np.zeros((8, 8), np.float32)), buf) # TypeError
read_geotiff(buf) # AttributeError
```

Suggested fix

Accept a binary file-like (anything with read/seek/tell, plus write for the writer) wherever the API takes a path. Add a _BytesIOSource class alongside _FileSource / _HttpSource / _FsspecSource and route _open_source to it for non-string inputs. For the writer, when the destination is file-like, build the TIFF into a BytesIO (or directly into the caller's buffer) and skip the os.replace/atomic-rename code path.

Out of scope:

• COG output to BytesIO (still needs full materialisation; would work but should be documented).
• VRT output to BytesIO (VRT is inherently a file-system construct that references sibling files).