[TRTLLM-12374][fix] Tighten DSv4 cache constraint floor to match warmup shapes

[lancelly]

Summary

DSv4-Pro hits cuMemCreate-FAIL during V2 KV cache initialization at small batch sizes (e.g., max_batch_size=8), with accumulated allocation reaching ~124.7 GiB even though the user-derived quota_from_max_tokens is only 117.84 GiB. This is
not a batch-size problem — the constraint floor is over-reserving.

Root cause

_storage_manager.py::_compute_slot_count_for_level takes:

quota = max(min_quota_from_constraints, user_quota)
                                       
so once min_quota_from_constraints exceeds user_quota, the user's free_gpu_memory_fraction is silently overridden.                                                                                                                                      
                                                    
The previous Constraint 1 was:                                                                                                                                                                                                                          
                                        
KVCacheDesc(capacity=max_seq_len, history_length=0)                                                                                                                                                                                                     
                                                                                                                                                                                                                                                        
history_length=0 represents the freshest state of a request that has the full max_seq_len capacity — a worst case that never actually occurs at runtime. With 64-layer DSv4-Pro and max_seq_len=300K, every pool group's min_slots was forced to        
max_seq_len / tokens_per_block ≈ 2344, including SWA and SSM pools that should collapse to their windowed working set. The resulting floor (~125 GiB) directly exceeded the user quota and triggered the OOM.                                           
                                                    
Fix                                                                                                                                                                                                                                                     
                                                                                                                                                                                                                                                        
Reshape the constraints to match the two real warmup shapes used by _capture_generation_cuda_graphs and _general_warmup_impl:                                                                                                                           
                                                                                                                                                                                                                                                        
┌────────────┬────────────────────────────────────────────────────────────────────────────────────────────────┬────────────────────────────────────────────────────┐                                                                                    
│ Constraint │                                          Warmup form                                           │                    KVCacheDesc                     │
├────────────┼────────────────────────────────────────────────────────────────────────────────────────────────┼────────────────────────────────────────────────────┤
│ C1         │ CUDA graph generation warmup — one decode request at the tail of max_seq_len                   │ capacity=max_seq_len, history_length=max_seq_len-1 │                                                                                    
├────────────┼────────────────────────────────────────────────────────────────────────────────────────────────┼────────────────────────────────────────────────────┤
│ C2         │ General / chunked-prefill warmup — one fresh context request of the per-iteration token budget │ capacity=max_num_tokens, history_length=0          │                                                                                    
└────────────┴────────────────────────────────────────────────────────────────────────────────────────────────┴────────────────────────────────────────────────────┘

[jiaganc]

LGTM