Fix relative import resolution

[gcv]

This PR improves name resolution for relative using/import with leading dots (., .., ...) and forward references (imports that appear before the target module is defined in the same file tree). It also adds a diagnostic when the number of leading dots exceeds the available module nesting.

Today, LanguageServer/StaticLint resolves relative imports only when the target module is already bound, and often treats top-level file scopes as “parents”, which makes sibling imports fragile. Common project layouts (one wrapper file including multiple modules) hit:

• using ..Sibling before Sibling is defined → unresolved references in the importing module
• import ..Another (no selectors) → module name “Another” is not bound locally, so qualified calls (Another.f()) don’t resolve
• leading dots that exceed the root silently do nothing (no diagnostic)

This PR implements:

• Late retry (forward-relative imports): when a relative segment can’t be resolved yet, schedule the import expression and its enclosing module for the existing ResolveOnly pass; in that late pass, re-run import resolution before resolving references. This lets sibling/grandparent modules that are defined later in the file tree become visible for references in the importing module.
• Resolve from module Scope: allow resolving a module’s own name when iterating scope.modules that contain a module Scope (not only ModuleStore), so identifiers like Sibling bind to the module in analysis.
• Bind bare import names: plain import Module (no selectors) now binds the module’s identifier in the current scope (matches Julia semantics), so qualified calls like Another.f() resolve without needing selectors.
• New lint code RelativeImportTooManyDots with message: “Relative import has more leading dots than available module nesting.”

Implementation details

• imports.jl:
  • resolve_import_block: walk leading dots up from state.scope; set RelativeImportTooManyDots when exceeding the root; when a segment is unresolved (cand === nothing), push the import expr and enclosing module expr into state.resolveonly; return to retry later.
  • resolve_import: selector-head failure (root2 === nothing) also schedules a retry.
  • _mark_import_arg: when usinged == false (plain import), bind the imported module name into scope.names; using keeps existing behavior (add module to scope.modules).
  • add_to_imported_modules: ensure scope.modules initialized when first used.
• references.jl:
  • resolve_ref_from_module(::Scope): resolve the module’s own name (bindingof(scope.expr)) and exported names via scope_exports.
• StaticLint.jl:
  • ResolveOnly: call resolve_import(x, state) before resolve_ref/traverse (idempotent; only updates unresolved imports).
• linting/checks.jl:
  • add RelativeImportTooManyDots to LintCodes and descriptions.

Tests

• Forward relative using/import: module C using ..Sibling binds Sibling; calls like Sibling.g() resolve even if Sibling is defined later in B.
• Bare import binds name: import ..Another makes Another visible, so Another.f() resolves.
• “Too many dots” lint: import ....X triggers RelativeImportTooManyDots.

Compatibility and performance

• No breaking changes; behavior now matches Julia semantics more closely.
• Work is gated: late import resolution only runs in the ResolveOnly pass (which exists today).
• Minimal additional cost; only schedules retries when an import segment can’t be resolved yet.