feat(drivers): expose authenticated RPC caller to @rpc handlers

[cagataycali]

Summary

Adds get_rpc_source_device() so device drivers can perform per-call authorization without changing their handler signatures.

The @rpc wrapper already receives the authenticated source device id (injected by DeviceRuntime). This change additionally publishes it on a contextvar for the duration of the handler call and resets it afterwards, so a handler can look up who called it and gate state-mutating operations accordingly.

Motivation

Device drivers that wrap physical hardware need to authorize callers on state-mutating RPCs (e.g. execute/stop). Today the source identity is only used for logging/tracing and is popped before the handler body runs. Exposing it via a small, opt-in accessor lets driver authors add caller-based access control as a security-hardening measure.

Backward compatibility

Fully backward compatible:

• Handlers that don't call get_rpc_source_device() are unaffected.
• Returns None when no source identity is present (local routine/internal calls).
• source_device is still consumed by the wrapper (not leaked into handler kwargs).

Changes

• drivers/decorators.py: _rpc_source_device contextvar + get_rpc_source_device() accessor; set/reset around the handler invocation.
• drivers/__init__.py: export get_rpc_source_device.
• tests/test_rpc_source_device.py: visible-inside-handler, None-when-absent, reset-after-call, and kwarg-non-leak.

Tests

tests/test_drivers.py .......................................... (60 passed)
tests/test_rpc_source_device.py .... (4 passed)

[kavya-chennoju]

Approving — tested end to end.

E2E (real D2D, Zenoh mesh): ran two live DeviceRuntime devices (controller-1 + sensor-001); controller-1 called invoke_remote("sensor-001", "whoami"). The runtime stamped source_device itself, and the handler read it back via get_rpc_source_device():

[sensor-001] whoami handler sees source_device = 'controller-1'
PASS: handler saw caller='controller-1' (expected 'controller-1')

Also confirmed the absent path: an agent-tools caller (not a device) yields caller=None, as intended. Unit tests pass (54).

The contextvar plumbing, reset-after-call, and no-kwarg-leak all behave correctly. LGTM as a caller-identity hook.

Non-blocking note for a follow-up: source_device is currently self-asserted by the caller and not yet bound to a transport-authenticated identity, so it shouldn't be relied on for authorization decisions until receiver-side identity verification lands. Tracking that separately.