Migrate to Pydantic v2

[r4victor]

Closes #1844

The PR is mostly a mechanical migration from Pydantic v1 features to Pydantic v2 features as described in https://pydantic.dev/docs/validation/dev/get-started/migration/. Parsing and serialization of all models is preserved with a few exceptions noted below. Python 3.14 is now supported for dstack installation and in run configurations. Overall, the PR improves validation and serialization speed, CLI startup times (due to pydantic-duality drop), code maintainability and drops several dependencies.

Main changes:

• pydantic-duality replaced with extra="forbid" CoreModel + per-call extra="ignore". Every model was previously generated as a request/response pair by generate_dual_core_model, where the X.__response__ variant allowed unknown fields. Pydantic since version 2.12 supports per-call extra= argument, so this is how now dstack does the validation in two different modes without pydantic-duality.
• __root__ replaced with RootModel.
• Types with custom parsing implementing __get_validators__ replaced with types implementing __get_pydantic_core_schema__ (Duration, Memory, Range, etc).
• Fixed annotated stored types and json input types. (In some cases, the annotated type was different from the actual stored type or a json type was missing.) Defined proper json input types on types instead of fields (e.g. Duration) and introduced OptionalOffableDuration and OptionalIdleDuration types to avoid duplicating per-model/per-field validators and json input types. Improves configurations json schemas and references.
• Dropped orjson and replaced CustomORJSONResponse with Pydantic v2 native serialization.
• @validator => field_validator, @root_validator => model_validator, GenericModel => BaseModel, Generic[T], model *() methods => model_*() methods.

Behavior changes:

• UTC datetimes now serialize as Z, not +00:00. This is Pydantic v2 default behavior that I think is not worth overriding. The dstack clients (Python, Go, JS) should all handle Z fine. Some HTTP API client may break (e.g. Python's fromisoformat() before 3.11) but this should be uncommon and is accepted.
• Fractional floats are rejected where an int is declared. nodes: 2.5 was allowed and truncated. Now it is a validation error; nodes: 2.0 still parses as 2.

Breaking changes:

• dstack Python API now works with Pydantic V2 models.
• dstack plugins now work with Pydantic V2 models.