Descriptor cache lookups can be reduced

[marvin-j97]

In the worst case, a table lookup can require 3 hits:

• filter
• index
• data

When any block is accessed, the file could maybe be stack allocated for the remainder of the lookup and reused for the other lookups. Reduces cache lookups to 1, always.

But it requires remodelling how load_block is used a little. We may need a &mut ReadContext or whatever.

Then, benchmark changes in a fully cached workload.

[marvin-j97]

The savings here are minimal I think - it requires that all blocks are uncached. The three cache accesses are miniscule compared to the I/O times. So maybe this could be explored in the future, but I don't think it makes too much difference in an I/O bound workload.