Expose explicit sandbox backend modes

[seanwevans]

Motivation

• Make the isolation backend an explicit API choice so callers cannot confuse an execution cell (subinterpreter) with stronger OS-level boundaries (process / microvm).
• Fail closed for boundary modes that are not yet implemented so callers can opt into explicit blast-radius semantics without silently changing behavior.

Description

• Add a BackendMode type and constants DEFAULT_BACKEND, SUPPORTED_BACKENDS, and IMPLEMENTED_BACKENDS and validation helpers _normalize_backend and _require_implemented_backend in pyisolate/supervisor.py.
• Thread a backend: BackendMode parameter through the public Supervisor.spawn(...) and module-level spawn(...) helpers, pre-validating and fail-closing on unimplemented backends, and store the chosen backend on the sandbox thread as ._backend with a Sandbox.backend property accessor.
• Wire backend through high-level helpers by adding a backend parameter to the @sandbox(...) decorator and to Pipeline.add_stage(...) and use the value when spawning stages in pyisolate/sdk.py.
• Export the backend symbols from the package pyisolate (pyisolate/__init__.py) and update the docs (README.md, API.md, docs/execution-model.md) to clearly distinguish backend="subinterpreter", backend="process", and backend="microvm" and to note that only subinterpreter is implemented in this build.
• Add unit tests in tests/test_supervisor.py that assert backend="subinterpreter" is selectable, that process / microvm fail closed with NotImplementedError, and that unknown backend values raise ValueError.

Testing

• Ran type/compile checks with python -m py_compile pyisolate/supervisor.py pyisolate/sdk.py pyisolate/__init__.py, which succeeded.
• Applied formatting with black on the modified files, which completed successfully.
• Ran targeted unit tests: pytest tests/test_supervisor.py and pytest tests/test_sdk.py tests/test_supervisor.py, and those test runs passed (all tests in those runs succeeded).
• Ran the full test suite (pytest) and observed many unrelated failures caused by an existing test stub altering the BPFManager.load signature during metrics tests; these failures are not caused by the backend plumbing and indicate a separate test-environment stub interaction that must be addressed independently.

---

Codex Task