Report measured host RAM capacity progress

[PaulFidika]

Implements python-gen-worker tracker #548.

Host-RAM admission failures now carry typed required/observed byte counts and a process-scoped generation. After an explicit owner/pin release, the worker reports HOST_CAPACITY_PROGRESS only when newly measured headroom actually increases enough to satisfy the remembered requirement. No elapsed-time retry, lifecycle timeout, environment knob, or broad CUDA cache flush is introduced.

Reconnect handling retains active failures and only undelivered satisfying progress. Its bounded queue preserves active-failure priority, replaces stale same-ref generations, fences blocked stale full-state producers, and clears progress on the same successful stream-write boundary already used for durable results. Old protocol-v3 consumers continue to parse failures and ignore the additive progress enum/fields.

Validation:

• full filtered suite: 1,137 passed, 13 skipped, 1 known clean-master CUDA-only lane deselected
• focused transport/lifecycle suite: 63 passed
• repeated stale-generation/reconnect/cancellation stress
• mypy (115 files), Ruff, build, HTTP timeout guard
• generated protobuf output verified deterministic
• independent exact-head review approved commit 58548bb87e4d616241cae97acd77ea7cfc8d9ad2