Previously, if we copied the last byte for a length of 100, we'd
recalculate the read span 100 times and memmove one byte 100 times,
which resulted in a lot of overhead.
Now, if we know that we have two consecutive copies of the data, we just
extend the distance to cover both copies, which halves the number of
times that we recalculate the span and actually call memmove.
This takes the running time of the attached benchmark case from 150ms
down to 15ms.
This error was introduced by 9a7accdd and had a significant impact on
`BufferedFile` behavior. Hence, we started seeing crash in test262.
By itself, the issue was a wrong calculation of the internal reading
spans when using the `read` and `until` parameters. Which can lead to
at worse crash in VERIFY and at least weird behaviors as missed needles
or detections out of bounds.
It was also accompanied by an erroneous test.
This patch fixes the bug, the test and also provides more tests.
This parameter allows to start searching after an offset. For example,
to resume a search.
It is unfortunately a breaking change in API so this patch also modifies
one user and one test.
The class is very similar to `CircularDuplexStream` in its behavior.
Main differences are that `CircularBuffer`:
- does not inherit from `AK::Stream`
- uses `ErrorOr` for its API
- is heap allocated (and OOM-Safe)
This patch also add some tests.
