## Description
When viewing photos/videos and enabling fullscreen (with a tap on
scree), the bottom bar wasn't disappearing like it used to. This PR
fixes this regression.
This should _significantly_ (hours => seconds) improve the performance
of the metadata parsing step during import of large Google Takeout zips,
and bring them to par as if the user had drag-and-dropped the unzipped
folder instead.
In my monkey tests, it seems that beyond a certain zip size (few GBs),
reopening the handle to a stream zip overshadows the time taken to read
the individual JSONs. This effect seems to grow very big for big zips to
a point where the metadata parsing step takes hours.
But note that I'm only testing this on synthetic exemplars I've created.
After merging it'll also need testing on more realistic huge takeout
examples.
## Description
- Use cached results for magic section so that it does't anymore have to
wait for ML framework to be initialised and doesn't have to re-compute
results every time, which means faster loading of the search tab.
- For internal users, all results in
[here](https://discover.ente.io/v1.json) will show up.
- For non-internal users, once available, results will be limited to 4.
- 4 random prompts are selected from
[here](https://discover.ente.io/v1.json) with non-empty results and are
cached.
- The cache updates when the data updates
[here](https://discover.ente.io/v1.json) (checks size to compare) or in
3 days since the last update.
```lib/core/error-reporting/super_logging.dart:20:8: Error: Error when reading 'lib/services/machine_learning/face_ml/face_ml_exceptions.dart': No such file or directory
import "package:photos/services/machine_learning/face_ml/face_ml_exceptions.dart";
^
lib/core/error-reporting/super_logging.dart:239:20: Error: 'CouldNotRetrieveAnyFileData' isn't a type.
if (error is CouldNotRetrieveAnyFileData ||
^^^^^^^^^^^^^^^^^^^^^^^^^^^
Target kernel_snapshot failed: Exception
FAILURE: Build failed with an exception.
