No scores when candidates parameter is added

Nobody has claimed this yet.

No scores are returned when you provide the `candidates` parameter for KeyBERT()

```
from keybert import KeyBERT

doc = """
Kos. Griekenland staat bekend om de prachtige eilanden waar je terecht kan voor zonovergoten vakanties.
"""
kw_model = KeyBERT()
keywords = kw_model.extract_keywords(doc, candidates=['Griekenland', 'Kos'])
```
Shows the warning message:

```
\venv\lib\site-packages\sklearn\feature_extraction\text.py:1369: UserWarning: Upper case characters found in vocabulary while 'lowercase' is True. These entries will not be matched with any documents
warnings.warn(
```
and keywords variable is returned empty.

Without the candidates paramater it does return a result with scores:
`keywords = kw_model.extract_keywords(doc)`
Result:
`[('griekenland', 0.5619), ('zonovergoten', 0.5024), ('bekend', 0.4398), ('prachtige', 0.4118), ('terecht', 0.4039)]`

When I change the candidates words to lower case words or when I add `lowercase=False` to the CountVectorizer it seems to return the words with a score as expected.:

`keywords = kw_model.extract_keywords(doc, candidates=['griekenland', 'kos'])`

In version 0.6.0 of KeyBERT() it wasn't an issue if the candidates words where capitalized.

```
count = CountVectorizer(
ngram_range=keyphrase_ngram_range,
stop_words=stop_words,
min_df=min_df,
vocabulary=candidates,
**lowercase=False**
).fit(docs)
```

Strangely enough it does seem to work on one of the virtual environments I've been using for a while, but I can't get it to work on newly installed environments even when I replicate it with the same versions of the packages installed. I expected the bug was in one of the installed packages, but this does not seem the case.

Start at KeyBERT.extract_keywords and inspect how the candidates parameter is passed to scikit-learn's CountVectorizer, especially its vocabulary and lowercase handling. Run the provided reproduction with capitalized candidates such as 'Griekenland' and 'Kos'; done means those candidates return scores without the uppercase-vocabulary warning, while existing lowercase behavior remains intact.

Written by the indexing model from the issue text.