Split harp.device into core, client and schema

[glopesdev]

harp.device becomes an empty PEP 420 namespace so that a device package can be harp.device.<name>, sitting beside the library portions rather than in a separate harp.devices tree that differs by one character.

harp is already a namespace, since no distribution ships harp/__init__.py. This does the same one level down. The requirement is that nothing ships harp/device/__init__.py, which harp-device did until now.

The three portions

harp/device/core/     the core register set every device carries
harp/device/client/   talking to a device
harp/device/schema/   building a device interface from a device.yml

core holds the 15 core registers, their 4 payload classes, 4 enums and REGISTER_MAP. client holds Device, Subscription, EventHandler, ITransport, TransportError and HarpFramer. schema holds create_device_module, parse_device_schema, DeviceModule, DeviceModuleLike and ConverterContext.

The split follows the dependency graph rather than cutting across it. core imports nothing from the other two, and client and schema each depend only on core, so there are no cycles and no module had to be divided.

harp-device remains a single distribution shipping all three portions. Nothing changes at the packaging level beyond dropping the __init__.py.

Why the names

core is the term this ecosystem already uses for the register set in core.yml, so harp.device.core.WhoAmI reads as it should. That is also the reason the device runtime did not go there. Putting Device in core would have read as though it were part of the core register set.

client matches how the reference material already describes these implementations when comparing pyharp against Bonsai.Harp.

Notable details

_schema/ flattened into schema/, and _emit_module.py became schema/_module.py. The three test modules that reached into the old harp.device._schema aggregate now name the owning module directly, which is more precise than the aggregate was.

docs/api/device.md had to be rewritten, and mkdocstrings failing on harp.device.Device is what caught it. Sphinx-style cross-references in docstrings were swept as well, since those go stale silently.

test_common_registers_are_not_device_module now asserts against harp.device.core, which is the more precise statement of what it always meant.