DocumentSplitter treats multi-character separators as regex (wrong splits, re.error crash on "++")

Description

DocumentSplitter passes its separator to Series.str.split() without regex=False, so pandas treats multi-character separators as regular expressions:

• DocumentSplitter(separator="..") on text a..b..c produces segments ['', '', '', 'c'] (regex .. matches any two chars) instead of ['a', 'b', 'c']
• separator="||" splits on empty string alternation — wrong boundaries
• separator="++" crashes with re.error: nothing to repeat at position 0
• the documented round-trip with DocumentJoiner is broken, because the joiner always joins with the literal separator

Steps to reproduce

import pandas as pd
from nemo_curator.stages.text.modules import DocumentSplitter

# crash case
DocumentSplitter(separator="++")
# re.error: nothing to repeat at position 0  (raised per batch, at str.split)

# silent corruption case
pd.DataFrame({"text": ["a..b..c"]})["text"].str.split("..")
# [['', '', '', 'c']]

Expected behavior

The separator is documented as a plain string ("The separator to split the documents on") and DocumentJoiner treats it as literal — splitting should too.

Actual behavior

df[self.text_field].str.split(self.separator) — pandas only treats single-char string patterns as literal; multi-char patterns go through re.

Environment

nemo-curator main (93b48525), Python 3.12

Start at DocumentSplitter in nemo_curator.stages.text.modules and inspect the pandas Series.str.split call using self.separator. Reproduce the documented cases for "..", "||", and "++", then verify splitting treats the separator literally and preserves the DocumentJoiner round-trip behavior. Add or update regression coverage for these cases and confirm the crash and incorrect boundaries are gone.