Protenix bond merging changes crossed copy pairings

## Description

`ProtenixInput.to_json()` can collapse crossed copy-specific covalent bonds into a copy-agnostic bond. Loading that JSON expands it as identity-paired bonds, so the round trip changes molecular topology.

## Minimal reproduction

Create two entities with `count=2` and these bonds at the same residue/atom positions:

```text
entity 1 copy 1 -> entity 2 copy 2
entity 1 copy 2 -> entity 2 copy 1
```

Then run:

```python
encoded = source.to_json()
roundtrip = ProtenixInput.from_json(encoded)
```

Observed result at current `main` (`72117242154d0c9f52ce727dc352210aab76d495`):

```text
before: [(1, 2), (2, 1)]
JSON: one bond with copy1/copy2 omitted
after: [(1, 1), (2, 2)]
```

Expected: serialization should preserve `[(1, 2), (2, 1)]`.

## Root cause

`_merge_covalent_bonds()` groups by entity, position and atom while ignoring copy IDs. It removes both copy fields whenever the number of grouped bonds equals both entity counts. `from_json()` interprets missing copy fields as identity pairing (`c -> c`), which is not equivalent for crossed or otherwise permuted pairings.

## Impact

A valid-looking Protenix JSON round trip can silently connect different molecular copies than the input requested.

## Why this issue does not include a PR yet

A safe fix needs the maintainers to confirm the intended Protenix serialization contract: whether non-identity copy pairings must always remain explicit, and which complete copy-pair sets are legal to compress. Preserving every copy field is lossless but may give up schema-level compression expected by downstream Protenix consumers. Once that contract is confirmed, a focused round-trip regression test and minimal merge predicate change should be straightforward.

## Environment

- Python 3.11.15
- PXMeter current `main`: `72117242154d0c9f52ce727dc352210aab76d495`
- Reproduced entirely in memory; no model, CIF or external service required

Start with ProtenixInput.to_json(), from_json(), and _merge_covalent_bonds(), then run the in-memory crossed-pair reproduction from the issue. Confirm with maintainers which copy-pair sets may omit copy fields; done means the agreed serialization contract preserves [(1, 2), (2, 1)] after a JSON round trip and includes a regression test.

Written by the indexing model from the issue text.