We switched from tab-separated to comma-separated starting with that version

# Conflicts:
#   rust/README.md

It needs to be the relative path from the prefix rather than the absolute URL

This happens to a handful of origins, due to how the dataset export works:
#4795

Instead, this skips origins with no known URLs and tries to find another one

Which itself switches to pytest-postgresql 5.0, which changes the object
the 'dsn' attribute is on.

Related to swh/meta#5075.

3min -> 2min on the prod graph

The Java backend does not provide labels for these endpoints, so these
tests are not very useful for that backend.

And run it on `example-labelled` and `example-transposed-labelled`
(though the former was already generated using webgraph's own
`build_eliasfano` command)

This does not seem to have an impact, but it can't hurt

I only tested on a small bit vector and assumed longer ones would have the
same structure, but they don't.

Java writes them in small chunks, separated by 5 bytes each, which makes
it impossible to use with NumberMmap.

Instead, LongArrayBitVector will now load the whole file in memory like
the Java class instead of mmapping. We should probably change this in the
future, to either mmap it and work around the weird format, or change
the on-disk format so LongArrayBitVector isn't needed anymore.

It kills the session-scoped gRPC server, causing any test running after itself
to fail. This currently does not happen because it is last in alphabetical
order; but a future commit will change that.