terraform: Stabilize the variable_validation_crossref experiment

[apparentlymart]

In #34947 we introduced a language experiment that would permit variable validation rules to refer to other objects declared in the same module as the variable.

Now that experiment is concluded and its behavior is available for all modules.

This closes #25609.

---

This final version deviates slightly from the experiment: we learned from the experimental implementation that we accidentally made the terraform validate command able to validate constant-valued input variables in child modules
despite the usual rule that input variables are unknown during validation, because the previous compromise bypassed the main expression evaluator and built its own evaluation context directly.

Even though that behavior was not intended, it's a useful behavior that is protected by our compatibility promises and so this commit includes a slightly hacky emulation of that behavior, in eval_variable.go, that fetches the variable value in the same way the old implementation would have and then modifies the HCL evaluation context to include that value, while preserving anything else that our standard evaluation context builder put in there.

That narrowly preserves the old behavior for expressions that compare the variable value directly to a constant, while treating all other references (which were previously totally invalid) in the standard way. This quirk was already covered by the existing test TestContext2Validate_variableCustomValidationsFail, which fails if the special workaround is removed.

[apparentlymart]

I've prepared this today in the hope that the experiment feedback is positive and we choose to move forward with stabilization, because it's easier to write a stabilization changeset with the experience of implementing the experiment still fresh in mind.

However, I'm going to leave this in draft until we've actually included the experiment in at least one alpha release and gotten some satisfying feedback about it. If feedback suggests changing approach then this I'll close this PR and replace it with another that implements another round of the experiment.