The caches decided when to collect from the driver's process-wide usage
estimate, which varied by about 4x between reads of the same moment on this
setup and is not what the allocator is holding. Read VMA's own accounting
instead, and let the collector stand down on the same figure it stood up on
rather than a stale tally.
Adds a small deferral guard so a completion callback cannot re-enter the
resource caches while the caller still holds their state.