When a segment qualified for compaction (had < 50% live entries or data)
but its live_size exceeded max_size (or NumLive exceeded max_count), and
no segments have been accumulated yet, the function returned
{[], All} where All still contains the problematic segment.

compaction: fix infinite loop

The current approoch of reading the snapshot index then the live
indexes is prone to a race condition where the snapshot state
is updated in betweeen reading the snapshot index and the live indexes.

E.g.

Snapshot is taken at index 10 with [7, 5, 3, 1]
Snapshot is taken at index 15 with [13, 11, 5, 3, 1]

Segment writer reads live indexes [7, 5, 3, 1] but snapshot index 15.

13 and 11 will not be written but 7 will be unnecesarily.

Segment writer: read the full snapshot state ETS record atomically.

ra_log_segment:close/1 in append mode now propagates errors from
sync and file:close instead of ignoring them. The segment writer
asserts close succeeds when flushing mem tables and defers pivot
file deletion until after the successor segment is opened.

Report errors on segment close instead of silently discarding them