feat: production-ready kernel execution — replace all ecosystem stubs with real dispatch

[mivertowski]

Core infrastructure:

• Add TypeErasedBatchKernel<K,I,O> wrapper for type-erased batch execution via JSON serde
• Add TypeErasedRingKernel<K,M,R> wrapper for type-erased ring message dispatch
• Add register_batch_typed() for zero-boilerplate factory registration with type inference
• Add register_batch_metadata_from() and register_ring_metadata_from() helpers
• Add execute_batch() convenience method on KernelRegistry
• Export new types through prelude (BatchKernelDyn, RingKernelDyn, TypeErased* wrappers)

Ecosystem — real kernel execution replacing all placeholder stubs:

• Axum: execute_kernel() now creates kernel instances from registry factories, serializes
  input JSON to bytes, calls execute_dyn(), deserializes output — with timeout support
  and proper error differentiation (batch vs ring vs not-found)
• Tower: KernelService::execute() uses same factory-based dispatch pattern
• gRPC: KernelGrpcServer::execute_kernel() with timeout and DEADLINE_EXCEEDED handling
• Actix: KernelActor handler bridges async execute_dyn() via block_in_place
• Ring kernels return informative errors guiding users to Ring protocol/gRPC streaming

Health & observability improvements:

• Health endpoint aggregates registry status and error rate into component health
• Prometheus metrics endpoint reports per-domain kernel counts, batch/ring breakdown,
  error rate, and request duration alongside existing counters

Domain crate registrations (all 14 crates, 106 kernels):

• Migrate from register_metadata() to factory-based registration throughout
• Batch kernels with BatchKernel<I,O> impl use register_batch_typed() for full execution
• Batch kernels with GpuKernel-only use register_batch_metadata_from() for discovery
• Ring kernels use register_ring_metadata_from() for metadata registration

All 895 tests pass. Zero clippy warnings.

https://claude.ai/code/session_01CwVtdXfd1urE8Ab6He77c5