We've ended up implementing something approximately like this in a few
places now, so this is a centralized version that we can consolidate on
moving forward, gradually removing that duplication.

Previously the "providers" package contained only a type for representing
the schema of a particular object within a provider, and the terraform
package had the responsibility of aggregating many of those together to
describe the entire surface area of a provider.

Here we move what was previously terraform.ProviderSchema to instead be
providers.Schemas, retaining its existing API otherwise, and leave behind
a type alias to allow us to gradually update other references over time.

We've gradually been shrinking down the responsibilities of the
"terraform" package to just representing the graph components and
behaviors anyway, but the specific motivation for doing this _now_ is to
allow for other packages to both be called by the terraform package _and_
work with provider schemas at the same time, without creating a package
dependency cycle: instead, these other packages can just import the
"providers" package and not need to import the "terraform" package at all.

For now this does still leave the responsibility for _building_ a
providers.Schemas object over in the "terraform" package, because it's
currently doing that as part of some larger work that isn't easily
separable, and so reorganizing that would be a more involved and riskier
change than just moving the existing type elsewhere.

Our existing functionality for dealing with references generally only has
to concern itself with one level of references at a time, and only within
one module, because we use it to draw a dependency graph which then ends
up reflecting the broader context.

However, there are some situations where it's handy to be able to ask
questions about the indirect contributions to a particular expression in
the configuration, particularly for additional hints in the user interface
where we're just providing some extra context rather than changing
behavior.

This new "globalref" package therefore aims to be the home for algorithms
for use-cases like this. It introduces its own special "Reference" type
that wraps addrs.Reference to annotate it also with the usually-implied
context about where the references would be evaluated.

With that building block we can therefore ask questions whose answers
might involve discussing references in multiple packages at once, such as
"which resources directly or indirectly contribute to this expression?",
including indirect hops through input variables or output values which
would therefore change the evaluation context.

The current implementations of this are around mapping references onto the
static configuration expressions that they refer to, which is a pretty
broad and conservative approach that unfortunately therefore loses
accuracy when confronted with complex expressions that might take dynamic
actions on the contents of an object. My hunch is that this'll be good
enough to get some initial small use-cases solved, though there's plenty
room for improvement in accuracy.

It's somewhat ironic that this sort of "what is this value built from?"
question is the use-case I had in mind when I designed the "marks" feature
in cty, yet we've ended up putting it to an unexpected but still valid
use in Terraform for sensitivity analysis and our currently handling of
that isn't really tight enough to permit other concurrent uses of marks
for other use-cases. I expect we can address that later and so maybe we'll
try for a more accurate version of these analyses at a later date, but my
hunch is that this'll be good enough for us to still get some good use out
of it in the near future, particular related to helping understand where
unknown values came from and in tailoring our refresh results in plan
output to deemphasize detected changes that couldn't possibly have
contributed to the proposed plan.