fix(engine): apply shrinking cache resizes before growing ones in live rebuild (Closes #643)

[void-mckenzie]

Closes #643.

Engine.rebuild_runtime_cache applied its three resizes in a fixed MoE -> KV -> mamba order
regardless of direction, and each pool frees only its own old tensors, so a mixed-direction
resize allocated the larger MoE cache while the old KV/mamba pools were still resident.
validate_rebuild prices only the target geometry, so a plan that passed the fit check could
OOM mid-rebuild (and the failed rollback after such an OOM is #526).

The rebuild now compares each pool family's old-vs-new footprint (new
BaseKVCachePool.rebuild_footprint classmethod reusing the kv_cost plumbing
validate_rebuild already uses; DSV4 overrides it with _dsv4_pool_sizes +
dsv4_pool_bytes, since its kv_cost takes no num_swa_pages) and runs the shrinking
resizes before the growing ones, preserving MoE -> KV -> mamba order within each phase.
Each pool still frees its own tensors first, so after the shrink phase the resident total is
sum(min(old, new)) and the grow phase rises monotonically to sum(new): the transient
peak never exceeds max(current, target), both of which already fit. All-grow and
all-shrink requests keep today's behavior.

Tests

New tests/engine/test_rebuild_runtime_cache_ordering.py: the issue's scenario (fake pools
under a shared cap, moe 40/kv 50 -> 60/30, cap 100) raises torch.OutOfMemoryError from the
MoE-first resize on unfixed main after the fit check passes, and completes after the fix with
peak == target total and the KV realloc ordered before the MoE realloc; plus reverse-direction
and rollback-restore guards. test_every_kv_pool_answers_the_sizing_surface gains the hook.

tests/engine/ + tests/kvcache/test_kv_cache_rebuild.py + tests/kvcache/test_pool_sizing_surface.py:
136 passed, 2 pre-existing env-gated skips. DSV4 / hybrid-SWA / V4.1 pool suites: 47 passed.

Hardware

RTX 5090, nvidia/Qwen3.8-Flash-Next-NVFP4 (offload MoE, moe 3400 slots / kv 317440 tok /
1.97 GiB free after CUDA graphs), command ft ctl cache rebuild --moe 4300 --kv 160k:

• main @ 0781324: passes the fit check, OOMs in the MoE grow (420 MiB request, 16 MiB
  free); rollback is rejected and the server latches failed (fix(engine): ft ctl cache live resize: OOM during rebuild breaks rollback and wedges the server #526's wedge).
• this branch: identical boot and command -> last rebuild ok, pools at moe 4300 /
  kv 163840, generation verified after the live resize.

The mid-rebuild-OOM rollback failure itself remains #526.