P2P shuffle is slow with string dtypes

```python
import coiled
import dask.dataframe as dd
from dask.distributed import wait

cluster = coiled.Cluster(
n_workers=30,
worker_cpu=4,
region="us-east-2", # start workers close to data to minimize costs
arm=True,
)

client = cluster.get_client()
```

```python
# this takes 1m21s
df = dd.read_parquet("s3://coiled-datasets/uber-lyft-tlc/")
df = df.set_index("request_datetime", shuffle="tasks").persist()
_ = wait(df)
```

```python
# this takes 2m12s
df = dd.read_parquet("s3://coiled-datasets/uber-lyft-tlc/")
df = df.set_index("request_datetime", shuffle="p2p").persist()
_ = wait(df)
```

GIL contention is very high during the p2p shuffle (also during tasks) and cpu usage is at 100+%, implying, maybe, that the creation/deletion of lots of Python objects is slowing us down considerably.

cc @hendrikmakait @jrbourbeau