GeoTIFF: centralize VRT capability validation (#2321 sub-task 2)

[brendancol]

Sub-task of #2321.

VRT capability checks are scattered across read_vrt (in _backends/vrt.py) and the open_geotiff .vrt branch, and some checks fire only once a chunk is actually decoded. That means a malformed VRT can build a dask graph successfully and then blow up inside a chunk function, far from the call site.

Add a single validator that runs against an already-parsed VRTDataset before any read work starts. It should cover:

• CRS compatibility
• dtype compatibility across bands
• band count sanity
• nodata policy
• transform orientation (flips the read path does not honour)
• pixel-size compatibility
• SrcRect vs DstRect sanity
• the supported resampling set

Both read_vrt and open_geotiff('foo.vrt') should call the validator so the two entry points produce the same rejections. Raise a typed VRTUnsupportedError (a subclass of the existing geotiff error class in _errors.py) and put the offending source path and field name in the message.

The validator must accept already-parsed structures. Do not re-parse the VRT XML inside it.

Acceptance

• Unsupported VRT features raise at graph build or eager-read setup, not deep in a chunk function.
• read_vrt(bad.vrt) and open_geotiff('bad.vrt') produce the same error type and message for the same input.
• One negative test per validator rule.