The `collapseIPv6Number` function was incorrectly producing IPv6 strings
with multiple "::" segments in certain cases. This was due to a flawed
replacement logic that could introduce a "::" and then later identify
another segment as compressible.

The function has been rewritten to follow the standard algorithm for IPv6
address collapsing (RFC 5952):
1. Expand the input IPv6 address to its full 8 hexadecatets.
2. Identify all sequences of zero-value hexadecatets.
3. Select the longest sequence (if multiple of same longest length, select the first one).
4. If this sequence consists of two or more hexadecatets, replace it with "::".
5. Otherwise (if the longest sequence is just one hexadecatet, or no zeros), no "::" compression is applied unless the entire address is zeros (which becomes "::").
6. All non-compressed hexadecatets are shortened by removing leading zeros.

This change resolves issue #99.

Added comprehensive test cases to `spec/IPv6UtilsTest.ts` to cover:
- The specific failing cases reported in the GitHub issue.
- Addresses with multiple zero sequences of varying lengths and positions.
- Addresses starting or ending with zero sequences.
- The all-zero address.
- Addresses with no compressible zeros (single zero blocks).
- Idempotency for already collapsed addresses.
- Addresses with prefixes.

All tests, including the new ones (total 322), pass with these changes.