This commit adds a memory monitor to the kv fetcher infrastructure
that's initialized by its users. When kv fetches occur, the new
infrastructure ensure that there's always at least 1 kilobyte allocated
for the fetch before it happens. Once the fetch returns, the accounting
is adjusted to include the entire size of the fetch. Subsequent fetches
that return less memory do *not* ratchet down the allocation size to
preserve safety and reduce some pointless allocation adjustment
churning.

This memory monitoring is gated behind a new cluster setting called
`sql.scan_memory_accounting.enabled` in 20.1. In 20.2, this memory
monitoring will always be on, and no such cluster setting will exist.

Release note (sql change): the memory used by disk scans is accounted
for, reducing the likelihood of out of memory conditions that result in
process crashes as opposed to SQL out of memory error messages. This new
behavior is off by default, and is gated behind the
sql.scan_memory_accounting.enabled cluster setting.

Release justification: fixes for high-priority bugs in existing functionality