Migrate common.ai provider to pydantic-ai 2.x and remove the <2 cap

[kaxil]

Migrates apache-airflow-providers-common-ai to pydantic-ai 2.x and removes the temporary <2 cap added in #68933.

pydantic-ai 2.x removed the instrument keyword from the Agent() and Agent.from_file() constructors; it is now only the agent.instrument property. PydanticAIHook.create_agent forwarded a caller-supplied instrument straight into the constructor, so create_agent(..., instrument=...) raised TypeError: got an unexpected keyword argument 'instrument' on 2.x. The hook now pops a caller-supplied instrument out of the constructor kwargs and applies it through the property, so a caller still overrides the provider's auto-instrumentation.

Changes

• Instrument handling (hooks/pydantic_ai.py): pop instrument before constructing the agent, then apply it via the agent.instrument property.
• Instrumentation format v4 to v5 (observability.py): formats 2 to 4 are deprecated in 2.x and emit a warning; v5 is the current default.
• Type annotations: Agent[None, ...] becomes Agent[object, ...] in the hook overloads and operators, matching the 2.x generic-deps default (object, not None). Type-checking only; runtime unchanged.
• Dependency floor: pydantic-ai-slim>=1.99.0,<2 becomes >=2.0.0; re-locked (pydantic-ai-slim and pydantic-graph 1.107.0 to 2.1.0).

Behavior changes (from pydantic-ai 2.x, surfaced through the provider)

• Telemetry: agent-run spans now report token usage under gen_ai.aggregated_usage.*, while per-model-call spans keep gen_ai.usage.*. This avoids double-counting in backends that sum a parent span and its children. Documented in observability.rst and asserted in the end-to-end span test. Dashboards or alerts reading run-level usage from gen_ai.usage.* should switch to gen_ai.aggregated_usage.*. Tracing is opt-in ([common.ai] otel_export_enabled) and off by default.
• end_strategy default flipped from early to graceful: function tools requested in the same response as a successful output tool now run instead of being skipped.

Design notes

• The floor is raised to >=2.0.0 rather than keeping 1.x/2.x dual support: format v5 does not exist on 1.x, dual-version branches would be untestable in CI (the lock resolves a single version), and the provider is pre-1.0. Raising a third-party floor is allowed; the release-manager-only rule applies to apache-airflow-providers-* dependencies.

Known follow-up

• The code-mode and skills extras still carry their pre-2.x adapter floors (pydantic-ai-harness>=0.3.0, pydantic-ai-skills>=0.11.0). The 2.x-native pydantic-ai-harness (0.5.0, which requires pydantic-ai-slim>=2.1.0) currently falls outside the repository's exclude-newer window, so raising those extra floors is left as a follow-up once it resolves.

Closes #69122