[Python] Add alpha TypedDict support

[lmazuel]

Add a model mode "typeddict" that would support some basic scenarios (at least for now)

Suppported:

• Input only for TypedDict, output returns raw JSON is fine
• Basic types (boolean, int, string)
• Enum (generate the enum as usual, and put the type in the TypedDict)
• Union (assuming all parts of the Union are in the supported list)
• List and Dict
• Operation treats that as if it receives a dict, and assume it's the correct JSON and send it to the pipeline

Not supported (refuse to generate)

• Discriminator
• Readonly
• Datetime
• Bytes
• Extends record (additional properties)
• Anything else that I don't list, since for now we really want a alpha support :)

CC @johanste

[iscai-msft]

Look into output deserialization, it might work. we're more concerned with input though

[johanste]

Input is the most valuable use of TypeDict. But input vs. output/deserialization only comes into play for operations (methods/functions) though, correct? So less about the generation of a TypeDict and more about how we generate operations using said types?

[msyyc]

One concern: when Typespec introduces unsupported type for TypedDict model mode with evolution, existing SDK users who used TypedDict before may meet generation error then have to use other model mode which may confuse users and bring breakings.

[johanste]

Can you demonstrate how that would be breaking? We already support mappings (JSON) as inputs. A TypedDict simply tells you the schema of said mapping.

[msyyc]

Can you demonstrate how that would be breaking? We already support mappings (JSON) as inputs. A TypedDict simply tells you the schema of said mapping.

In this issue's description #8800 (comment), there are unsupported type Anything else that I don't list, since for now we really want a alpha support :) for TypedDict. Since some of them are quite common (e.g. Datetime), Python emitter will have to let user change to another model when Typespec user introduce Datetime unsupported for TypedDict in new version?

Update: after discussion with @iscai-msft, python emitter will report diagnostics for unsupported type in TypedDict models mode. And I have no questions about this feature for now.

[johanste]

Typed dicts supports all datatypes in json. As such, date-time is simply a str in the typed dict (assuming it is iso8609 encoded - it would be an int if it was ticks etc.). Exactly like it is in the Mapping view of the model.

While it is possible to dig yourself into a hole when it comes to changes over time, I don't think it is much different than for our existing models. And you have to work pretty hard in order to do so with typed dicts.

[johanste]

Not supported (refuse to generate)

We can refuse to generate for now (since the services we are interested in doesn't have those problems. Well, maybe we swapped out closed tagged unions for our discriminator extensible pattern. Need to look at that). But per my comment above, all of those types are available "raw" through __getitem__ on the model. Putting it another way, the TypedDict really just represents the typed version of __getitem__ on the dict-like view of a model.

additionalProperties with well-known property names (with values of a different type) would be the biggest outlier. But that is an anti-pattern that we try really hard to avoid so having to cast to something else doesn't seem to be a big deal.