Parsing a regex with groups that may match empty strings seems to be incorrect sometimes

### Description

Consider a regex of the form `a(b|)c`. This should match `ac` and `abc`. Some regexes of this form seem to behave incorrectly. To try to make a slightly more concrete or realistic example, one might want to search some logs for something like `[a-z-]*(|-prod|-main) `.

I cobbled together some tests against a recent version of the lucene repo. I have three docs which respectively contain:
- `foo-bar-baz`
- `foo--baz`
- `foo-test-baz`

And then I spell a regex to match only the first two docs a bunch of different ways:

1. `.*foo-(bar|)-baz.*` - IllegalArgumentException: expected ')' at position 18
2. `.*foo-(|bar)-baz.*` - 0 matches but 2 expected
3. `.*(foo-(bar|)-baz).*` - IllegalArgumentException: expected ')' at position 20
4. `.*(foo-(|bar)-baz).*` - 0 matches but 2 expected
5. `.*foo-(bar|())-baz.*` - 2 matches
6. `.*foo-(bar|()?)-baz.*` - 2 matches
7. `.*foo-(bar|#?)-baz.*` - 2 matches
8. `.*(foo-(bar|())-baz).*` - 2 matches
9. `.*(foo-(bar|()?)-baz).*` - 2 matches
10. `.*(foo-(bar|#?)-baz).*` - 2 matches
11. `.*foo-(bar)?-baz.*` - 2 matches

The first four cases seem incorrect to me.

I came to this investigation after some problems with elasticsearch (v8.12.2, using lucene 9.9.2) where regexes following pattern number 5 also failed. Maybe that is some useful context.

### Version and environment details

I added tests by modifying `lucene/core/src/test/org/apache/lucene/search/TestRegexpQuery.java` with a base revision of `cd1a4ecc9ead8e06b08f3bc2016297525f65b37c`.

Here are other details in case they are relevant:
OS: linux (el8 with a 6.12.39 kernel)
Java: openjdk version "24.0.1" 2025-04-15
This is on an x86_64 box.

Start with lucene/core/src/test/org/apache/lucene/search/TestRegexpQuery.java and reproduce the first four patterns against the three documents. Trace the regexp parsing entry point used by those tests, then verify that the empty-alternative forms produce the expected matches while the existing successful cases continue to pass.

Written by the indexing model from the issue text.