…id windowed-pool deadlock

When KVCacheManagerV2 is built with a typical_batch describing the
working set (e.g., max_batch_size concurrent decode requests with
capacity=max_seq_len), windowed pool groups whose window_size is
smaller than tokens_per_block previously collapsed to min_slots=1
because get_stale_range() consumed all but one block per request, and
_compute_min_slots_from_constraints() only enforced an absolute floor
of 1 slot per pool group. With more than 1 concurrent decode request,
the V2 scheduler could not find a free slot in windowed pools and
deadlocked.

Synthesize a constraint from the typical_batch: one
KVCacheDesc(capacity=tokens_per_block, history_length=tokens_per_block-1)
per request. For every pool group, this yields non_stale=1 per
request, so the new floor is len(typical_batch.kv_caches) — large
enough to support the scheduler's full concurrency.

Signed-off-by: tensorrt-cicd <90828364+tensorrt-cicd@users.noreply.github.com>