NUMERIC reads: full digits arrive, the decode drops them

The read-side follow-up promised in crate/sqlalchemy-cratedb#300, with the measurement #652 was missing. As @matriv said there, the full number reaches the client; the digits are dropped by this driver's decode.

On CrateDB 6.4.2, with `1.234567890123456789012345` stored in a `NUMERIC(38, 24)` column, the raw `/_sql` response carries every digit:

```json
{"cols":["n"],"rows":[[1.234567890123456789012345]],"rowcount":1,"duration":105.078}
```

`cursor.fetchone()[0]` on the same query returns `1.2345678901234567`, a float. The response is decoded with `orjson.loads` ([http.py line 244](https://github.com/crate/crate-python/blob/a61a3e5f30fafce3956b800a3d6f9a256aa0fac1/src/crate/client/http.py#L244)), which parses every JSON number to a float64 and has no `parse_float` hook; stdlib `json.loads(raw, parse_float=Decimal)` on those bytes returns `Decimal('1.234567890123456789012345')`.

Writes are already exact, since `Decimal` goes out as a string (#751). A stdlib decode with `parse_float=Decimal` would make reads match, but it changes the returned type for every float column and gives up orjson's speed, so it likely wants to be a connection option rather than the default.

Start in src/crate/client/http.py at line 244 and reproduce the NUMERIC example against CrateDB 6.4.2. Compare the current orjson decode with stdlib json.loads using parse_float=Decimal, then determine how a connection option should expose exact numeric reads while preserving existing float behavior; done means the reported digits are retained and the option is covered by verification.

Rédigé par le modèle d'indexation à partir du texte de l'issue.