AlgoScan

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

Previously, we always used 2 bytes to encode branch targets, now we
use a varint encoding so that

1. Short jumps take only two bytes total
2. Even jumps across very large programs (currently illegal) would be possible.

We currently encode branch labels in two bytes. That wastes space for
small jumps, and prevents ever jumping more than 32k (which is ok for
now, but we're contemplating large programs).

This commit introduces varint encoded labels for branches. But it
doesn't us them to very good effect yet.  It always encodes in two
bytes, because the assembly changes are tricky -- since we don't know
the offset of targets in the first pass, we don't know how much room
the instruction will eventually take. We'll need a pass that adjust
the sizing of branch intructions while keeping track of the movement
of labels.

Instead of calling replaceBytes once per ref and then walking all stored
positions each time, collect all changes first, sort them, and apply them
in one left-to-right scan that builds the new program bytes and a cumDelta
array simultaneously. cumDelta[p] records the total size shift from all
replacements whose replaced region ends at or before p, so any stored
position can be adjusted with a single array lookup in any order.

Complexity goes from O(k*n + k*m) to O(n + k + m), where k = number of
refs, n = program size, m = total stored positions.

Deploy `main` to production

merge main-dist into main