Memory leak in BERTopic.transform

Nobody has claimed this yet.

The following minimal example repeatedly calls `BERTopic.transform` on some random texts, and it records the memory usage.

```python3
# memleak_bertopic.py

import random
import string

import psutil
from bertopic import BERTopic

def random_strings() -> [str]:
return [''.join(random.choices(string.printable, k=500)) for _ in range(200)]

model = BERTopic()
model.fit(random_strings())

print("iteration,memory_usage_in_MiB", flush=True)
for iteration in range(99999999999999):
model.transform(random_strings())
memory_usage_in_MiB = psutil.Process().memory_info().rss / (1024 * 1024)
print(f"{iteration},{memory_usage_in_MiB:.2f}", flush=True)
```

Output:

```
iteration,memory_usage_in_MiB
0,759.21
1,796.00
2,765.11
3,801.29
[...]
99,986.20
100,971.39
101,988.67
102,1006.21
[...]
488,2099.64
489,2103.87
490,2108.17
491,2121.31
[...]
761,2843.27
762,2844.84
763,2827.00
764,2859.39
[...]
961,3486.86
962,3481.87
963,3493.12
964,3511.06
[...]
```

The problem can be reproduced by building the following `Dockerfile` (no GPU):

```Dockerfile
FROM python:3.10.9
RUN pip install bertopic==0.12.0 psutil==5.9.4
ADD ./memleak_bertopic.py /
RUN python /memleak_bertopic.py
```

Start by running the provided memleak_bertopic.py example using the Dockerfile and confirm that repeated BERTopic.transform calls increase RSS. Then trace transform's memory behavior within BERTopic and compare it across iterations. Done means the reproduction no longer shows unbounded memory growth under the supplied workload.

Written by the indexing model from the issue text.