fix(pydantic): support pydantic-ai v2 in restate.ext.pydantic while keeping v1

[selimacerbas]

Problem

restate.ext.pydantic is written against pydantic-ai v1 and crashes on import under
pydantic-ai-slim==2.0.0:

>>> from restate.ext.pydantic import RestateAgent
ModuleNotFoundError: No module named 'pydantic_ai.builtin_tools'

The pydantic_ai optional dependency is declared open-ended (pydantic-ai-slim>=1.68.0),
so 2.0.0 resolves and then fails at import time.

What changed (v1 1.68+ AND v2 2.0.0 both supported)

v1 → v2 change	Handling (kept v1-compatible)
pydantic_ai.builtin_tools deleted → pydantic_ai.native_tools; AbstractBuiltinTool → AbstractNativeTool, tools.BuiltinToolFunc → tools.NativeToolFunc	try the v2 names, except ImportError fall back to the v1 module aliased to the same local names. (v1 1.68+ already exposes the new names, so no deprecation warnings on v1.)
pydantic_ai.mcp.MCPServer renamed to pydantic_ai.mcp.MCPToolset — separate classes: 1.x concrete servers subclass MCPServer, 2.0 only has MCPToolset	Detect against the tuple of MCP base classes present in the installed version (MCP_TOOLSET_CLASSES); isinstance accepts a tuple. A single aliased name would silently stop wrapping MCP servers on whichever line it didn't match.
Agent.run(builtin_tools=…) and its **_deprecated_kwargs shim removed in v2; native tools now go via capabilities=[NativeTool(tool), …]	RestateAgent.run translates builtin_tools → capabilities=[NativeTool(t) …] when the NativeTool capability is available (1.68+), and never forwards builtin_tools on v2; oldest 1.x falls back to forwarding builtin_tools=.

The RestateAgent(event_stream_handler=…) public API is unchanged. The handler was only
ever passed at run time (still accepted in v2) and stored on RestateModelWrapper, never
to a pydantic-ai constructor — so RestateModelWrapper.request_stream continues to read
self._event_stream_handler exactly as before. (For reference, v2's
capabilities.ProcessEventStream(handler=…) is the new way to register a handler on the
agent, and its docstring confirms the run-time event_stream_handler parameter still
observes live events under durable-execution integrations — so no change is needed here.)

Repro

uv add restate-sdk 'pydantic-ai-slim[openai,mcp]==2.0.0'
python -c "from restate.ext.pydantic import RestateAgent"   # before: ModuleNotFoundError

The break is iterative (builtin_tools module → BuiltinToolFunc → MCPServer), plus a
runtime-only TypeError on the first agent.run() from forwarding builtin_tools= to a
base run() that no longer accepts it.

Tests

Verified on both pydantic-ai-slim==1.107.0 and ==2.0.0 against this branch:

• import restate.ext.pydantic (clean, no deprecation warnings on v1)
• RestateAgent(inner, event_stream_handler=handler) construction; handler flows into
  RestateModelWrapper._event_stream_handler and the event_stream_handler property
• RestateAgent.run(builtin_tools=[WebSearchTool()]) translates to
  capabilities=[NativeTool(...)] and forwards no unknown kwargs (validated against the
  real base run() signature)
• A real MCP server (MCPServerStdio on v1, MCPToolset on v2) is wrapped as
  RestateMCPServer

ruff clean; pyright introduces no new errors (removes the v2-incompatibility ones).

There is currently no test fixture for restate.ext.pydantic in the repo — happy to
contribute the two smoke tests above + a pydantic-ai 2.x CI matrix entry as a follow-up if
that's useful.

Follow-up to #201.