This replaces the generator-as-a-couroutine with proper full async and uses rendercanvas's existing loop for scheduling. Throttling behavior is much more natural now and IO never blocks rendering even when waiting for Futures to resolve. The entire pipeline from getting the raw data --> window_funcs --> spatial_func is async and fully concurrent.

closes #1042

@apasarkar this should allow concurrent running of window_funcs and spatial_funcs written using torch that stay on the device until the very end, ex: torch.mean() for rolling mean windows instead of np.mean.

• need to benchmark and verify that the performance is just as good as before, especially with cuda arrays from masknmf.
• NDWSubplot.subplot property @apasarkar
• NDGraphic._set_indices_() doesn't take any arguments, it retrieves the latest indices directly from the ref_index instance. This is because since _set_indices_() is a scheduled request AND the indices dict is produced at the time it's scheduled in _render_indices(), the indices argument will not reflect the latest indices at execution time. So NDGraphic._set_indices_() just always grabs the latest indices when it executes.
• better x_range_mode logic in NDPostions, get rid of jittery and jigglyness when x_range_mode="auto", fix other bugs. _update_from_view_range() uses with block_indices_ctx() to any async scheduling of _set_indices_() when x_range_mode="auto" but we allow async scheduling to update all other NDGraphics. We then run_sync(self._set_indices_()) immediately so that the NDPositions graphic immediately reflects the latest polled camera state. This gets rid of the jitter and jiggly effect! 🥳
  • no x_range fighting when setting indices via _set_indices_
  • no x_range fighting with _update_from_view_range() - trying to figure this out
  • fix bug when width was 0 due to figure auto_scale when using VisibilitySelector and no lines are visible (i.e. initial selection is empty, blank subplot).
  • only one NDPositions LinearSelector per-subplot, so multiple timeseries can be overlaid in one subplot and they use the same linear selector.
• better PlotArea.x_range and y_range setters and getters.
• keep UI throttling for imgui slider, and also use Future cancellation in ReferenceIndex render request logic.
• No throttling for other forms of viewing data, ex: play button, stepping through frames, using the LinearSelector, we want to see every single datapoint/frame of data when doing fine-grained inspection like this.
• make sure new async works when slider is moved around very fast with multi-video decoding with fava - almost there
• similarly, make sure new async works when viewing masknmf.DemixingResults and moving the slider very fast.
• tested examples: masknmf single session, ephys, kcenia behavior, gerbils. WIP masknmf multi session, things wonky with timeseries.
• cleanup comments and docstrings

@BalzaniEdoardo