It's a false dependency. The result isn't used and there are no
relevant side-effects.

They each have two call sites, and the sequencing is almost identical in
both cases.

`Compiler::register_plugins()` calls `passes::register_plugins()`, which
calls `Compiler::dep_graph_future()`. This is the only way in which a
`passes` function calls a `Compiler` function.

This commit moves the `dep_graph_future()` call out of
`passes::register_plugins()` and into `Compiler::register_plugins()`,
which is a more sensible spot for it. This will delay the loading of the
dep graph slightly -- from the middle of plugin registration to the end
of plugin registration -- but plugin registration is fast enough
(especially compared to expansion) that the impact should be neglible.

`Compiler::compile()` is different to all the other `Compiler` methods
because it lacks a `Queries` entry. It only has one call site, which is
in a test that doesn't need its specific characteristics.

This patch removes `Compile::compile()` and replaces the call to it in
the test with a call to `Compile::codegen_and_link()`, which is similar
enough for the test's purposes.

The goal of this commit is to improve the effectiveness of pipelining,
because metadata generation must complete before dependent crates can
begin compiling.

Unfortunately it's not much of a win. The biggest speed-up I have seen
is 1.07x, vs 1.79x for the original pipelining implementation. And it
slightly slows down the common case when an error occurs during
analysis, because metadata generation now precedes analysis.

Currently three tests fail with the change, due to bounds violations.
These tests are currently `git rm`'d, so that subsequent tests can run.