Issue a warning if loadable_variables does not work for a given parser (VirtualTiff)

[scottyhq]

While I think the new parser interface is really slick, it took me a while to realize that the round trip test (https://virtualizarr.readthedocs.io/en/stable/custom_parsers.html#testing-your-new-parser) suggested in the docs does not currently work with the VirtualTiff parser:

from obstore.store import GCSStore
from virtual_tiff import VirtualTIFF
from obspec_utils.registry import ObjectStoreRegistry

import xarray.testing as xrt
import xarray as xr 
# for rioxarray/rasterio backend
import os 
os.environ['GS_NO_SIGN_REQUEST'] = 'YES'

bucket = "solus100pub"
file = "cec7_0_cm_p.tif"
url = f"gs://{bucket}/{file}"
store = GCSStore(bucket=bucket, skip_signature=True)
registry = ObjectStoreRegistry({f"gs://{bucket}": store})

parser = VirtualTIFF(ifd=0)
manifest_store = parser(url=url, registry=registry)

with (
    xr.open_dataset(manifest_store, engine="zarr", zarr_format=3, consolidated=False) as actual,
    xr.open_dataset(url, engine="rasterio") as expected,
):
    xrt.assert_identical(actual, expected)
# AssertionError: Left and right Dataset objects are not identical

At least one reason for this is that unlike the common 'rasterio' engine to load GeoTiffs directly into memory, VirtualTiff does not currently parse coordinates (virtual-zarr/virtual-tiff#55). So the passing loadable_variables=['x','y'] to open_virtual_dataset has no effect. Currently you have to do some manual tweaking of the virtualized format or adjust after loading to get the coordinates and spatial metadata to pass such a test.

# Raise a warning here since can't load 'x' and 'y' into memory ?
ds = open_virtual_dataset(
    url=url,
    registry=registry,
    parser=VirtualTIFF(ifd=4),
    loadable_variables=['x','y'] # This has no affect because VirtualTiff does not parse coordinates!
)
ds

In any case, I don't think an identical match is necessary for a lot of use cases, but though it could be nice to issue some warning from open_virtual_dataset if the requested loadable_variables aren't actually loaded?

[TomNicholas]

This doesn't really make sense - the loading step is separate from the parsing step...

it could be nice to issue some warning from open_virtual_dataset if the requested loadable_variables aren't actually loaded?

That makes more sense - we should definitely at least warn, if not ouright error if the user tries to load variables that don't exist!!!

The faulty logic must be in here

VirtualiZarr/virtualizarr/xarray.py

Line 528 in ba1f464

def replace_virtual_with_loadable_vars(