[BUG] cudf.Series str.replace performance appears slower than nvstrings

[Garfounkel]

Describe the bug
cudf.Series str.replace performance appears slower than nvstrings. Probably related to #4885.

Steps/Code to reproduce bug

import nvstrings
from cudf import Series
import string

corpus = [
     'this is the first document.',
     'this document is the second document.',
     'and this is the third one.',
     'is this the first document?',
]
n = 500

corpus_wide = [doc * n for doc in corpus]

d = nvstrings.to_device(corpus_wide)
s = Series(corpus_wide)

print("nvstrings:", flush=True)
%time d.replace('[{}]'.format(string.punctuation), '')  # 0.1s
# nvstrings:
# CPU times: user 84 ms, sys: 28 ms, total: 112 ms
# Wall time: 111 ms

print("cuDF:", flush=True)
%time s.str.replace('[{}]'.format(string.punctuation), '')  # 30s
# cuDF:
# CPU times: user 18.6 s, sys: 12.1 s, total: 30.8 s
# Wall time: 30.8 s

Expected behavior
Performance should be more closely aligned with nvstrings.

[kkraus14]

cc @davidwendt for if you have any ideas