Shuffle not raising exception when on does not exist

[fjetter]

Upstream issue #11160

from dask.datasets import timeseries
ddf = (
    timeseries()
    .shuffle(on='does_not_exist')
    .repartition(partition_size='100MB')
    .compute()
)

This should raise an exception already at construction time but instead it just computes the dataframe and seems to return the full result.
Particularly with other expressions combined, this can cause very weird behavior. For example

import dask.dataframe as dd
timeseries().to_parquet("data_test.parquet")

# reading back again
ddf = (
    dd.read_parquet("data_test.parquet")
    .shuffle(on='does_not_exist')
    .repartition(partition_size='100MB')
    .compute()
)

causes the issue that's been reported upstream but only if repartition and read_parquet is used in combination with the erroneous shuffle

[fjetter]

This is also faulty without dask-expr enabled. We'll have different exceptions depending on whether there is a parquet reader or not

ddf = (
    # dd.read_parquet("data_test.parquet")
    timeseries()
    .shuffle(on='does_not_exist')
    .repartition(partition_size='100MB')
    .compute()
)