db: consider adapting LoadBlockSema to limit concurrent bytes loaded

[jbowens]

Memory for in-flight reads is allocated before a block can be inserted into the block cache, so these reads do not drawn down from the block cache's memory. A spike in concurrent block loads can cause a spike in memory allocations. The LoadBlockSema semaphore used to limit the count of concurrent loads would probably be more effective as a limit on the concurrent bytes loaded. This might even be a better proxy for concurrent read i/os, since large reads may translate to many individual block i/os (especially as it pertains to cloud platforms concurrent iop budget calculations).

Jira issue: PEBBLE-311