(core) be careful when reassigning a doc to a worker it was on before

Summary:
Importing a .grist document is implemented in a somewhat clunky way, in a multi-worker setup.

 * First a random worker receives the upload, and updates Grist's various stores appropriately (database, redis, s3).
 * Then a random worker is assigned to serve the document.

If the worker serving the document fails, there is a chance the it will end up assigned to the worker that handled its upload. Currently the worker will misbehave in this case. This diff:

 * Ports a multi-worker test from test/home to run in test/s3, and adds a test simulating a bad scenario seen in the wild.
 * Fixes persistence of any existing document checksum in redis when a worker is assigned.
 * Adds a check when assigned a document to serve, and finding that document already cached locally. It isn't safe to rely only on the document checksum in redis, since that may have expired.
 * Explicitly claims the document on the uploading worker, so this situation becomes even less likely to arise.

Test Plan: added test

Reviewers: dsagal

Reviewed By: dsagal

Subscribers: dsagal

Differential Revision: https://phab.getgrist.com/D3305

app/server/lib/DocSnapshots.ts

@@ -163,6 +163,18 @@ export class DocSnapshotInventory implements IInventory {
     });
   }
 
+  /**
+   * Wipe local cached state of the inventory.
+   */
+  public async clear(key: string) {
+    await this._mutex.runExclusive(key, async() => {
+      await this._flush(key);
+      const fname = await this._getFilename(key);
+      // NOTE: fse.remove succeeds also when the file does not exist.
+      await fse.remove(fname);
+    });
+  }
+
   /**
    * Remove a set of snapshots from the inventory, and then flush to S3.
    */