Improve memory estimate for metadata handler

[travisdowns]

Version & Environment

Redpanda version: 22.2

What went wrong?

The fix implemented in #4804 is in most cases a larger overestimate of the memory needed (so we went from not-large-enough to a too-large estimate, which is at least safe).

Instead, we can implement a more dynamic memory semaphore approach using a two phases: an initial small estimate for the initial request processing, then a second acquisition of units at the point in processing where we know the exact size of the response and intermediate structures.

This will both avoid the overestimate problem and narrow the region of time where the full memory amount is requested, increasing parallelism when metadata requests are in progress.

JIRA Link: CORE-969