Large bokeh plots struggle to squeeze in and out of pydantic models

[ml-evs]

Since merging #1311, there's a bit of a performance regression on the NMR in situ block (datalab-org/datalab-app-plugin-insitu#65), it seems like we're running into bottlenecks in serializing some block types that have large bokeh plot data. We should probably trust the bokeh data and skip validation for these fields, by default.

[ml-evs]

Definitely OOM problem, with peak memory usage around ~10x the size of the bokeh plot due to repeated serialization.

Simple fix: don't repeatedly serialize, and hope the types get sorted out.

Longer term fix (may need to wait for #1285) --- do not do any validation of the bokeh plots, just pass them through

Perhaps useful suggestion: add some memory profiling to our pytest suite

[ml-evs]

Just FYI, the memory issue was a bit of a red herring here; gunicorn worker was being sigkilled on the demo server due to timeouts, but the default message is to report a possible OOM.