Fix deploying to Heroku OOM issues [ci skip]

[camsaul]

After fixing the metabase-deploy repo I thought I'd give the "Deploy to Heroku" button a test to make sure things were working correctly. To my horror, they were not; the application was constantly running out of memory when Liquibase attempted to parse our "huge" migrations file. It seems like this is not a new issue: #3360, but it has certainly gotten worse over time as the migrations file has grown.

I tried every trick in the book to get things to work correctly without having to change anything in Metabase itself. After a few hours of experimenting I have found a set of JVM options that seems to resolve the issue.

Corresponding PR for deploy repo: metabase/metabase-deploy#18

It seems like sooner or later we'll have to split our migrations file into two or more files to resolve these memory issues, but this seems to be a short-term fix that works as we'd expect

[camsaul]

Alternatively we can specify in the app.json file that we should default to a different dyno size. However the next dyno with a large amount of memory (1GB instead of 512MB) is standard-2x, which costs $50 a month. I experimented with defaulting to that dyno size, and everything works perfectly, but going from free to $50/mo. seems like something we should try to avoid if we can