Bulk upsert silently reverts concurrent writes (missing attributes are backfilled from a stale, non-transactional read)

### 👟 Reproduction steps 1. Have a collection with several optional attributes, e.g. `status` (string) and `score` (integer), and a document `D`. 2. From two server-side clients (API key), run **concurrently**: - Client A: bulk upsert `[{ "$id": "D", "score": 42 }]` (only `score`) - Client B: bulk upsert `[{ "$id": "D", "status": "FT" }]` (only `status`) 3. Repeat under concurrency (or with a batch large enough that the two requests overlap). Both requests return HTTP 200/201 and report the row as upserted. ### 👍 Expected behavior Since each payload contains **disjoint attribute sets**, both writes should survive: `D` ends up with `score = 42` **and** `status = "FT"`. This is what happens with two concurrent *single* `updateDocument` calls, which are safe (see analysis below). ### 👎 Actual Behavior Whichever bulk upsert commits last **reverts the other one's columns to the values they had before both requests started** — even though those columns were not in its payload, and even though the API reported success for both writes. There is no error, no warning; the lost write is silent. The root cause is that bulk upsert is not a column-targeted patch. In `Database::upsertDocumentsWithIncrease()` (utopia-php/database): 1. The old version of each document is read **outside the write transaction**: - in database 3.x (Appwrite 1.8.x): a per-document `getDocument()`, which is served by the **Redis document cache** (unbounded TTL), so the merge base can be arbitrarily stale for frequently-read documents; - in database 6.x (Appwrite 1.9.x) and current `main`: a batched `find()` executed once at the start of the call — a direct DB read, but still outside the write transaction, so it goes stale while earlier chunks of the same call (or concurrent writers) commit. 2. Every optional attribute missing from the payload is then backfilled from that stale read: ```php // Force matching optional parameter sets // Doesn't use decode as that intentionally skips …