Fixes#8520
If the websocket is closed not by us, we automatically try to reconnect with it, and initialize the protocol again. **Restoring state (execution contexts, attached visualizations) is not part of this PR**.
It's a part of making IDE work after hibernation (or LS crash).
# Important Notes
It required somewhat heavy refactoring:
1. I decided to use an existing implementation of reconnecting websocket. Replaced (later discovered by me) our implementation.
2. The LanguageServer class now handles both reconnecting and re-initializing - that make usage of it simpler (no more `Promise<LanguageServer>` - each method will just wait for (re)connection and initialization.
3. The stuff in `net` src's module was partially moved to shared's counterpart (with tests). Merged `exponentialBackoff` implementations, which also brought me to
4. Rewriting LS client, so it returns Result instead of throwing, what is closer our desired state, and allows us using exponentialBackoff method without any wrappers.
- Improved performance by batching simulatenous node edits, including metadata updates when dragging many selected nodes together.
- Updated Vue to new version, allowing us to use `defineModel`.
- Fixed#9161
- Unified all handling of auto-blur by making `useAutoBlur` cheap to register - all logic goes through a single window event handler.
- Combined all `ResizeObserver`s into one.
- Fixed the behaviour of repeated toast messages. Now only the latest compilation status is visible at any given time, and the errors disappear once compilation passes.
- Actually fixed broken interaction of node and visualization widths. There no longer is a style feedback loop and the visible node backdrop width no longer jumps or randomly fails to update.
- Closes#8179
# Important Notes
- ⚠️ These tests are currently *not run* on any CI workflow.
- There is some unused code for mocking the PM. This has been intentionally kept, as this may be useful in the future.
Note that this may be useful for testing the dashboard, however the dashboard is currently only tested in cloud mode
- that is, without the backend switcher, and with only the remote backend available. As such, currently it uses HTTP API mocks, and no PM mock.
