MMseqs short-sequence matching drops duplicate database hits

## Description

`calc_mmseqs_seq_identity()` records only the first database entity when a short query sequence has multiple exact matches. The Smith-Waterman implementation in the same module correctly emits every matching database ID.

## Reproduction

Use a short query sequence `AAA` and two database rows with the same sequence but different IDs (`A_1` and `B_2`). Call `calc_mmseqs_seq_identity()` with `min_seq_length=25`. Because all sequences take the short-sequence path, MMseqs2 does not need to produce any alignments.

Expected rows:

```text
query_id db_id similarity
Q_1 A_1 1.0
Q_1 B_2 1.0
```

Actual result contains only `Q_1 -> A_1`.

## Root cause

The short-sequence branch appends `id_in_db[0]` instead of extending the result with all IDs stored in `db_short_seq_to_id[seq]`.

## Impact

Downstream low-homology classification can miss a training PDB that shares the exact short sequence. If another entity provides the common PDB link, an entry can be incorrectly retained as low-homology, causing benchmark leakage.

## Environment

- PXMeter commit: `72117242154d0c9f52ce727dc352210aab76d495`
- Python 3.11 and 3.13 on Windows

A focused regression test can cover the short-sequence branch without invoking MMseqs2.

Start at calc_mmseqs_seq_identity() and inspect its short-sequence branch, especially how db_short_seq_to_id[seq] is used. Add a focused regression test using query AAA and database IDs A_1 and B_2; done means both exact-match rows are returned without invoking MMseqs2.

Written by the indexing model from the issue text.