Try gen2 resource classes for Circle executors

[dschuff]

These cost 20% more but are supposed to be up to 40% faster

[sbc100]

We already have #27020

[dschuff]

Setting aside the build-linux step then, let's look at the rest:
here's a list of the jobs with their p95 duration over the last 30 days.

Comparing the times in that list with the runtimes of these tests (I have a spreadsheet, sorry it's internal) for the top 10 or so tests showed that only one of them was enough faster to cost less. Several were enough slower that it made me wonder if we added more tests recently and that caused the slowdown?

edit: here's 24h: https://app.circleci.com/insights/github/emscripten-core/emscripten/workflows/build-test/jobs?branch=b1d8acd4-8701-4c76-8a87-7ff2320aa95f&reporting-window=last-24-hours

Here's the sheet: https://docs.google.com/spreadsheets/d/1j4UfUNK0ba76aXqITxdBdYbNQpH3TiuiTmF_TaKx9l0/edit?gid=0#gid=0
This is a small sample and is subject to noise (just one gen2 run), and also this compares against 24h P95 (rather than mean or median). Also I don't have a good sense for what the distribution of times is, so I don't know how many gen2 tests we'd have to do to have good confidence.
But none of the top 13 tests are cheaper with gen2. 3 of them are slower (10-25%).