Previous splitting logic simply put all the remainder in the last
tranche, which could make the last tranche run significantly more test
cases. We now change it so the remainder is evened out across tranches.
To make the CI indicative, we now starting tracking the flaky tests
found when running on Windows. As a starting point, rather than ignore
the windows CI entirely, we now identify there are cases where lnd can
be buggy when running in windows.
We should fix the tests in the future, otherwise the windows build
should be deleted.
This commit adds a new flag to shuffle all the test cases before running
them so tests which require lots of blocks to be mined are less likely
to be run in the same tranch.
The other benefit is this approach provides a more efficient way to
figure which tests are broken since all the differnet backends are
running different tranches in their builds, we can identify more failed
tests in one push.
With this commit we create a new function that returns system wide
unique ports by using a single file to keep track of previously used
ports. We'll want to use this everywhere whenever we need to listen on a
new, random port during unit or integration tests.
Because we now have a unique source, we don't need to apply the port
offset that was used for the different tranches of parallel running
integration tests before.
This commit adds a new build tag `integration` and removes the old tag
`rpctest` for clarity. Multiple unnecessary usages of `build !rpctest`
is also removed.
