Performance: `_coerce_field_names_to_aliases` rebuilds on each validation and can block streaming CPU

[keatingw]

SDK Version (required)
7.0.9

Describe the bug
UniversalBaseModel._coerce_field_names_to_aliases is a mode="before" model validator which makes it run for every nested model on each parsed object. Each call rebuilds name to alias maps which should be constant per class.

For v2 chat_stream this work runs multiple times per SSE event and can spend a large amount of parse time in this validator per event - I think this is the same class as the other caching fixes for _type_adapter_cache and _type_hints_cache in recent releases

Suggested fix

Cache per class, keyed like _type_adapter_cache:

• Compute once per class: the map of field names whose alias differs, plus the ambiguous-key set (key that is one field's alias and another's name)
• Fast paths - most models have no differing aliases and can be returned as is; if no field-name keys are present we can skip the dict copy
• parse_obj_as's per-call has_pydantic_aliases scan asks the same question and can reuse the cache

Related to this - with caching parsing will still stay materially heavier than pydantic-core native behavior since the Fern-generated models seem to lean on smart unions instead of using a discriminator Pydantic can see