Support Tiered in-memory + SSD cache for data

Having an inmemory + SSD cache for data files could be very useful in the lance API.

Heres how it might look like

ds = lance.dataset(                                                                                                                     
      "s3://bucket/nuscenes.lance",                                                                                                       
      storage_options={                                                                                                                   
          "ssd_cache_dir": "/mnt/nvme/lance_cache",                                                                        
          "ssd_cache_size_mb": 100_000,
           "max_memory_cache_mb":1000,                                                    
      }
                                                                                                                                   
  ) 

How it should work:

Two-tier cache modelled after the Velox caching system:

Remote GCS/S3                                                                                                                           
      │  fetch on miss                                                                                                                    
      ▼                                                                                                                                   
Memory cache (hot, bounded)                                                                                                             
      │  evict to SSD when memory threshold hit                                                                                           
      ▼                                                                                                                                   
SSD cache (warm, bounded)                                                                                                               
      │  evict LRU when SSD limit hit                                                                                                     
      ▼                                                  
 Discard   

Start by locating the Lance API path that handles storage_options and the data-file fetching path. Compare the requested behavior with the Velox-style two-tier model described here. Done means supporting bounded memory and SSD caches, promoting data on misses, evicting from memory to SSD, and applying LRU eviction at the SSD limit.