fix: Remove get_cached() and use get() everywhere for cross-worker visibility

[mattf]

Remove get_cached() entirely from the DistributionRegistry protocol and implementations. All callers now use get() which implements cache-then-DB fallback semantics.

The old get_cached() was a cache-only accessor that returned None when a routing table lookup missed the in-memory dict. In single-process deployments this was fine, but in multi-worker uvicorn deployments each process has its own cache that is only updated by _ensure_initialized() at startup, register(), update(), and delete() -- all within a single process. A cache miss in worker A after worker B created a resource would always return None instead of falling back to the DB where worker B had persisted the object.

This is the exact bug the register() override already worked around: its doc comment explicitly says 'use super().get() (DB read) rather than self.get_cached() so that in multi-worker deployments, where each process has its own in-memory cache, we always read the authoritative stored object'. The same logic applies to all callers.

get_all() already had TTL-based cache refresh on every call, and that's used for list endpoints. But the routing path (get_provider_impl()) and the metrics helper (_get_provider_id()) both called get_cached(), creating a silent failure mode where freshly created resources were invisible until the next TTL refresh or a restart.

By removing get_cached() and using get() everywhere (which checks cache first, falls back to DB on miss, and updates the cache), we get immediate cross-worker visibility for all operations with only a modest additional DB round-trip cost on first access per resource (and that access gets cached for subsequent calls).

Changes:

• Remove get_cached() from DistributionRegistry protocol
• Remove get_cached() from DiskDistributionRegistry (was raising NotImplementedError)
• Remove get_cached() from CachedDiskDistributionRegistry (was just a dict get)
• Update get_provider_impl() in common.py to use get() (was using get_cached())
• Update _get_provider_id() in vector_io.py to be async and use get() (was sync and using get_cached()) -- all 6 call sites are in async methods and already awaited
• Update test mocks and test assertions that referenced get_cached()

closes #5008