LOS: Reduce is_mapped calls during VO bit scan

[wks]

Like in find_last_non_zero_bit_in_metadata_bytes, we should only call is_mapped when we reached a new chunk because the granularity of MapState is per chunk.

[wks]

This optimization is orthogonal to #1337. While #1337 reduces the overhead of each individual is_mapped invocation, this PR reduces the number of calls of is_mapped for the LOS.

(Note that the linear searching algorithm for non-LOS spaces already take MapState's chunk granularity into consideration.)

I ran benches/mock_bench/internal_pointer.rs on several revisions, and here are the performance comparison. Numbers are the average iteration time, and the unit is nanosecond. 1337 and 1341 are PR numbers. 1337 and 1341 is a branch with both PRs merged.

space	master	1337	1341	1337 + 1341
normal	101.66	83.862	94.948	78.371
LOS	294.80	30.533	94.320	27.795

Note that PR 1341 should not affect normal spaces, and the performance improvements are likely to be noises.

With either or both of PR 1337 and PR 1341 applied, the LOS searches faster (for the same object size 16KB) than other spaces because it searches at page granularity. This is the expected behavior.