Inconsistent behavior in `getBounds`, `cameraForBounds`, `fitScreenCoordinates` and `fitBounds` with `globe` projection

Nobody has claimed this yet.

When `globe` projection is used the following snippets yield inconsistent behavior when compared to `mercator` projection:

### `getBounds` and `cameraForBounds`:
```
const c = map.cameraForBounds(map.getBounds());
map.easeTo(c);
```

https://user-images.githubusercontent.com/67509069/164701266-16ce6e51-6faf-490f-94aa-74045d3ee2e7.mp4

The returned camera transform should remain the same similar to when `mercator` is in use:

https://user-images.githubusercontent.com/67509069/164701298-847b9f91-30c4-4bbf-ac70-743d6d0fb1e5.mp4

### `fitScreenCoordinates`:
```
map.fitScreenCoordinates([0, 0], [map.transform.width, map.transform.height], map.getBearing())
```

https://user-images.githubusercontent.com/67509069/164707548-bb06e6a5-52b9-4d3e-8a61-da3280a18d26.mp4

### `fitBounds`:
```
map.fitBounds(map.getBounds());
```

https://user-images.githubusercontent.com/67509069/164712924-6dee4d71-c681-473b-9971-c36d609e994f.mp4

### `boxZoom`:

https://user-images.githubusercontent.com/67509069/164719186-e14b7ae6-423f-4053-84ea-2c03f90a1b5f.mp4

Start by reproducing the globe-projection cases for getBounds, cameraForBounds, fitScreenCoordinates, fitBounds, and boxZoom using the snippets and recordings in the issue, then trace their camera and projection handling. Done means these operations preserve the expected camera behavior consistently with mercator, with regression coverage for the reported cases.

Written by the indexing model from the issue text.