feat: declarative worker model residency (protocol v3)

[PaulFidika]

Summary

• hard-cut the worker scheduler wire contract to protocol v3
• replace ordinary model DOWNLOAD/LOAD/UNLOAD commands with full-replace ordered desired disk refs and exact hot endpoint instances
• report accepted desired generations separately from residency convergence
• let active RunJob work preempt background reconciliation, then resume the current desired state when idle
• retain ModelOp only for ADOPT_COMPILE_CACHE; reserve removed enum values 1-3
• package the matching worker as gen-worker 0.26.0

Contract

DesiredInstance reuses the existing RunJob binding path: function name plus the complete slot-to-immutable-ref map identifies the concrete runnable instance. Same-generation desired state is a valid retry/presigned-URL refresh; only lower generations are stale. Protocol field numbers match the coordinated Tensorhub schema.

This PR must land with the matching Tensorhub protocol-v3 controller change. It should not be merged or released independently.

Validation

• focused declarative/adoption/real-gRPC suites: 85 passed
• final focused rerun: 54 passed
• mypy: 114 source files, zero issues
• ruff: all changed Python files pass
• HTTP timeout lint: pass
• package build: gen_worker-0.26.0 sdist and wheel pass
• full suite: 1070 passed, 13 skipped, 1 known environment-sensitive CUDA failure

The remaining full-suite failure is tests/test_content_lanes.py::test_lane_swap_moves_only_the_exclusive_module. It reproduces in isolation on the current shared GPU: the synthetic 2.3 GiB budget expects acme/a to demote to RAM, but the installed runtime keeps it in VRAM. None of this PR's touched code is on that budgeting/content-lane path.