i#7860 load bal: Add load balancing to drmemtrace analyzer (#7878)

When running drmemtrace analyzers with dynamic scheduling (i.e.,
core-sharded with live scheduling instead of replay), we want our worker
threads to have relatively even "activity" as seen by a simulator for a
reasonable final virtual schedule. Atomic activity counts are
periodically updated and examined to accomplish this. If a worker's
activity reaches a specified ratio versus the slowest worker, it sleeps
until it is under the ratio. A new CLI flag --max_load_imbalance sets
the max ratio; the default is 2.5.

Adds a unit test, though it would likely become flaky if waiting were
required. Confirmed manually that there is waiting in the 3 fast workers
for the slow worker #0 who ends up with fewer instructions but within
the ratio (there are no idles in this test):
```
12: [analyzer] Worker 2 @3000 waiting for slowest 3 @1000
12: [analyzer] Worker 3 @13000 waiting for slowest 0 @1000
12: [analyzer] Worker 1 @14000 waiting for slowest 0 @1000
12: [analyzer] Worker 0 waited 0 times for load balancing
12: [analyzer] Worker 1 waited 71 times for load balancing
12: [analyzer] Worker 3 waited 71 times for load balancing
12: [analyzer] Worker 2 waited 71 times for load balancing
12: shard 0 saw 215000 instructions
12: shard 1 saw 530000 instructions
12: shard 2 saw 535000 instructions
12: shard 3 saw 530000 instructions
```

Tested on larger traces on 80 cores with a live max of 60 with target
ratios at 2.0, 2.5, and 3.0, which were precisely hit in two consecutive
runs each with the 100K check cadence settled on in the code here
(auto-raising to 1M or higher reduces accuracy with some ratios climbing
too high). The check_load_balance() function does show up in some
profiles as high as 3.3% but seems worth the cost. When reducing below
1.5, targets are hard to reach, and
the overhead does go up: check_load_balance() is 4.6% in a 1.25-target
run and 15% in a 1.0 run in these experiments, with the ratio not
dropping under 1.5: so 1.5 may be a practical limit, as noted in the
option description, for now: which should be fine for initial use cases.

Fixes #7860