Stop inferring schemas for serialized graphs

[dglazkov]

Now that I've used the new syntax for a little bit, I am realizing that trying to infer input/output schemas for graphs is mostly in the way. It introduces artifacts that I didn't expect. Here are a couple of examples:

• in the blank.ts graph, it is surprising to see that it produces a graph requiring text input and text output. This is not something that the developer of the graph specified.
• when creating an inlined graph to iterate by a map, the schema inferred for the output node appears in the output of the map run, and has to be filtered out.
• graphs that don't have input/output schemas specified has data in their inferred schemas that are not very useful, like titles "The schema to convert to schemish" where the use of Schemish is an implementation detail.

Let's remove that machinery.

@seefeldb WDYT?

[seefeldb]

Sure, and this shows a few underlying issues that we might want to tackle:

• right now the schema information added with the new syntax is lost on serialization, except for when its used in automatic schema generation. so we could keep automatic schema generation but turn off the default assumptions it makes. but even better might be to serialize schema information on the wires, then use that at runtime. Maybe we call describe on input and output (which gets that on-wire schema passed in) to create the schema on the fly?
• map outputting schema: ah, that's probably because output emits it, but maybe it shouldn't? or more broadly, it feels like schema should be out of band, not a regular wire. congruent with the first point. This would then also support using graphs in map that do have schemas defined, e.g. because one is reusing some existing graph there.
• blank graphs: i think there are a few places where graphs that use 0 inputs or produce 0 outputs trip us up.
• for blank.ts specifically, it looks like it has a text wire, so i think all you see here is that if there is no type associated, it's assumed to be a string input. we could make it fail if there is no type associated instead, but i think that's debatable.

So short term you could leave the automatic generation in, but remove drop all the wires that don't have an explicitly set title. But the real solution would be to start serializing schema information as first class citizen on wires and maybe also on nodes.