Performance for plural endpoints is suboptimal

Some background: we use Kinto as a synced per-user data store, using a bucket per app deployment/environment, and a collection per user. Users can often have thousands of records (most fairly small, a handful per collection fairly large).

We're using `postgresql` for the `storage` and `permission` backends (and `memcached` for the `cache` backend).

The performance of plural endpoints (e.g. `/v1/buckets/my-app-staging/collections/my-user/records` to get all records in a collection) in the current server implementation is a bit disappointing (ignoring caching).

I've profiled the Kinto server using Sentry, by adding `traces_sample_rate=1.0, _experiments={"profiles_sample_rate": 1.0}` to the `sentry_sdk.init()` call. While the SQL queries themselves take a bit of time, it's also spending a considerable amount of time in library functions.

## JSON deserialisation

Swapping out [`json.loads`](https://docs.python.org/3.8/library/json.html#json.loads) for [`msgspec.json.decode`](https://jcristharif.com/msgspec/api.html#msgspec.msgpack.decode) for SQLAlchemy's JSON deserialisation gives a substantial improvement:

```
Benchmark (json.loads): curl http://localhost:8888/v1/buckets/my-app-development/collections/my-user/records
Time (mean ± σ): 1.490 s ± 0.114 s [User: 0.006 s, System: 0.008 s]
Range (min … max): 1.327 s … 1.879 s 100 runs
```

```
Benchmark (msgspec.json.decode): curl http://localhost:8888/v1/buckets/my-app-development/collections/my-user/records
Time (mean ± σ): 1.267 s ± 0.052 s [User: 0.006 s, System: 0.007 s]
Range (min … max): 1.150 s … 1.428 s 100 runs
```

This improved the performance by ~18% for this collection (~3000 records).

Start at the plural records endpoint and SQLAlchemy's JSON deserialization path, then reproduce the reported benchmark for a collection of about 3,000 records. Compare json.loads with msgspec.json.decode and profile the remaining query and library costs. Done means the endpoint's performance is measurably improved without changing its response behavior.

Written by the indexing model from the issue text.