Add a ThreadLocal metrics aggregator

[Swatinem]

Thus far, this only works for count metrics, but we can extend it in the future as well.

I would still call this pretty WIP for now. Ideally, I might extract all this metrics business into its own crate, or move it into the SDK crate.

[Swatinem]

I finished implementing this for all the metrics types we are using, meaning counter, gauge and distribution-like.

The results running symbolicator-stress look like this:

# no metrics

Workload 0 (concurrency: 50): 1378551 operations, 91903.40 ops/s

# unmodified `cadence` metrics

Workload 0 (concurrency: 50): 801336 operations, 53422.40 ops/s

# this PR with the thread-local aggregator

Workload 0 (concurrency: 50): 1202184 operations, 80145.60 ops/s

So this is a 50% improvement over using cadence on its own. And this reduces the overhead of collecting metrics to ~13% if my math is right.

For all the "distribution-like" metrics, we currently keep all the individual values, but we send them as batches. We could instead calculate the distribution and percentiles locally and avoid that, whereby potentially improving things further.