fix(runtime): report expected PTY support while deferred startup is pending

Fixes #1623

What

With create_or_reuse(..., defer=True), DeferredSandboxSession.supports_pty() returned False until bring-up finished. The Agents SDK checks that method synchronously, in Shell.tools() and at the start of ExecCommandTool.run(). A session that was still pending therefore:

• built its shell tools without write_stdin, and
• ran the first exec_command in one-shot mode, so yield_time_ms acted as a hard timeout and no session ID came back.

This takes approach 1 from the issue: say up front what PTY support to expect.

• register_backend(..., supports_pty=False) gains a flag next to supports_bind_mounts, plus a backend_supports_pty() helper. The built-in docker backend is declared PTY-capable (DockerSandboxSession.supports_pty() returns True).
• DeferredSandboxSession takes supports_pty and reports it while pending. Once resolved it still delegates to the real session.
• create_or_reuse passes backend_supports_pty(backend_name) into the deferred session.

Custom backends default to False, so they behave as before unless they opt in. That avoids assuming every backend supports PTYs. The defer=False path is unchanged.

Tests

New tests in tests/test_deferred_session.py:

• a pending session on a PTY backend reports supports_pty() and Shell.tools() includes write_stdin
• a pending session on a backend without the flag stays on exec_command only
• the issue's reproduction as a test: the SDK's real UnixLocalSandboxClient behind a gated deferred bring-up. The first ExecCommandTool.run() returns Process running with session ID and no timeout.

Before the change, the first and third tests fail. After the change, uv run pytest tests/test_deferred_session.py passes (13 tests). The full suite shows the same 19 failures as main (they come from existing _start_child_runner(skills=...) and settings fakes), and the three new tests are the only difference. ruff format --check ., ruff check ., mypy strix/ and bandit -r strix/ -c pyproject.toml are clean.

AI assistance: this change was prepared with Claude (AI coding assistants) and tested as described above.