The encoded "JSON" wasn't actually JSON but the JS compatible object
creation. This patch makes it proper JSON and therefore should resolve
the current config parsing issues.

This patch drops the additional environment variables generated by
Kubernetes for service discovery. There happen to match some immich
configurations and result in a broken configuration.

BREAKING CHANGE: This commit require manual adjustments of the database.

```
DROP SCHEMA vectors CASCASE;
DROP EXTENSION vectors;

TRUNCATE TABLE smart_search;
TRUNCATE TABLE asset_faces CASCADE;
TRUNCATE TABLE person;

ALTER TABLE smart_search
ADD COLUMN IF NOT EXISTS embedding vector(512); -- need to change the number here if using a non-default CLIP model

ALTER TABLE asset_faces
ADD COLUMN IF NOT EXISTS embedding vector(512);

SET vectors.pgvector_compatibility=on;

CREATE INDEX IF NOT EXISTS clip_index ON smart_search
    USING hnsw (embedding vector_cosine_ops)
    WITH (ef_construction = 300, m = 16);

CREATE INDEX IF NOT EXISTS face_index ON asset_faces
    USING hnsw (embedding vector_cosine_ops)
    WITH (ef_construction = 300, m = 16);
```

Reference:
https://github.com/immich-app/immich/discussions/7310#discussioncomment-8559595

This patch reverts the changes to immich to upgrade to 1.96.0. The
pgvecto.rs upgrade turned out to be too complicated, it might be worth
to rebuild the immich setup using pg_vector or drop immich entirely.

From the current point of view pgvecto.rs is too immature and too
involved to maintain in a private setup using standard components like
postgresql clusters using zalando postgres operator.

With all preparation done, immich can now be upgraded to 1.94.1. It just
needs a custom database image and a bunch of config options that
hopefully never break…

This version needs certain database extensions that are implemented as
part of the database setup.