feat(BA-7353): add the kernel creation payload type and event conversion errors

[fregataa]

Summary

Groundwork split out of #13746 so the JSON cutover can be reviewed without the type definitions attached to it. Nothing changes on the wire here — the new types have no caller yet, and the existing serialize() / deserialize() msgpack path is untouched.

• Add KernelCreationInfo to events/event_types/kernel/types.py, beside the events that carry it. It is the contract with the manager, not a rendering of the agent's resource spec: it carries only what update_kernel_status_running() reads — container_id, kernel_host, the two repl ports, service_ports and occupied_devices.
• Describe a held device once. attached_devices and the per-slot occupancy used to describe the same devices from two sides, both keyed by device name and then by device, so a unit metered along two axes — a cuda device is, with cuda.device and cuda.shares — appeared twice with its attributes duplicated. OccupiedDevices.units is keyed by DeviceName and then DeviceId, and each OccupiedDevice carries every slot it supplies.
• Carry the allocation in both units it is expressed in. allocated is in the scheduler's — keyed by ResourceSlotName, and what slot_totals sums into the kernel's occupancy. processing_units and memory_size are the same allocation in the device's own, mirroring AbstractComputeDevice. They cannot share one mapping: smp is not a slot, and a GPU's mem is not the host memory slot of the same name, so folding them together would put both into the resource accounting.
• Raise EventPayloadEncodingError / EventPayloadDecodingError from AbstractEvent.to_message() / from_message() so a caller sees one event-layer error rather than a Pydantic one. An event deriving BackendAISchema maps ValidationError to a BackendAIError of its own, so both forms are funnelled in.

Every field is JSON-representable and round-trips unchanged: amounts stay Decimal, mappings are declared Mapping so a consumer editing a payload it received is caught by the type checker, and there are no field defaults — a producer states the whole payload.

A non-finite amount is rejected at decode. Infinity belongs to a resource policy with DefaultForUnspecified.UNLIMITED, not to an allocation — a device supplies a finite share of what it has, and alloc_map never produces one.

Notes for the reviewer

• Where the device fields come from. The plugins fill DeviceModelInfo.data freely — its declared ComputedDeviceCapacity has never matched what anyone writes. Both accelerator plugins in reach (the enterprise CUDA one and the mock) write exactly smp and mem, which are AbstractComputeDevice.processing_units / .memory_size projected onto the kernel's share; the intrinsic cpu plugin writes cores, which is the kernel's total core count copied onto every device and which nothing reads. Hence two named fields rather than an open mapping, and cores dropped.
• model_name is kept even though it is not part of the creation contract: the GPU usage stats aggregate the device models a kernel ran on.
• slot_totals is a plain property, not a computed_field: a computed field is written into the payload beside the allocations it is derived from, where nothing reads it — a receiver recomputes it — and it can disagree with the value next to it.
• The middle commits move the payload into a common/interchange/ package and then back out again. That package was the answer to "where does an inter-component payload live"; defining it beside its events made it unnecessary. Kept as separate commits so the review discussion stays readable.

Follow-up for #13746

The manager writes kernels.attached_devices as {"cuda": [{"model_name", "data": {"smp", "mem"}}]}, and the GPU usage stats read it back through that shape. Wiring this payload up has to reshape it, or those stats go to zero.

Test plan

• tests/unit/common/events/ — 41 passed
• pants test tests/unit/common/events::
• pants check :: — the new types are unused by production code in this PR, so type checking is the main guard

Resolves BA-7353

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[HyeockJinKim]

If possible, could you change it from ResourceSlot to ResourceSlotEntry wherever it appears?

[HyeockJinKim]

If possible, it would be best to manage the Dict by separating the types.
A three-level dictionary just doesn't seem to be very readable.

[HyeockJinKim]

I think “schema” might be a better fit than “data,” but even that’s a bit ambiguous.
In my opinion, it would be best to separate only the values that are actually needed so that the values stored in “data” within the Agent are distinct from those passed via the DTO. Could you expand the scope of this task a bit to address this?

[HyeockJinKim]

We don't provide information about sub-packages, though.

[HyeockJinKim]

In my opinion, the interchange package doesn't seem necessary; I think the best approach would be to define a separate type in the event section and implement the conversion logic there.

[HyeockJinKim]

On the event side, it looks like "ResourceSlotName" should be the key.