Add semantic lint diagnostics and unused-variable warnings

[julia-script]

Summary

Introduce non-blocking semantic lint diagnostics, starting with unused bindings and parameters, while keeping compilation gating restricted to errors.

Scope

• Extend compiler diagnostic severity beyond the current error-only model.
• Add a lint phase/category whose warnings do not make Diagnostic.hasErrors fail.
• Derive usage from semantic and ownership facts, not token spelling or raw occurrence count.
• Distinguish declaration, read, write, move, and drop contexts.
• Warn for unused locals, write-only bindings, and unused parameters under a documented policy.
• Define an intentional-unused convention if Silk needs one.
• Keep public/external-interface declarations out of workspace-only unused conclusions.
• Allow warnings without fixes.
• Produce compiler-owned source actions only when evaluation, effects, ownership, and cleanup are proven preserved.
• Establish configuration/suppression behavior without lint-specific compiler shortcuts.

Safety cases

• Removing let value = run operation() must not silently remove execution.
• Removing an affine binding must not change required cleanup.
• A write-only mutable binding is distinct from a never-referenced binding.
• Removing a parameter is a call-site refactor, not a local deletion.
• Same-spelled shadowed bindings must be analyzed independently.

Acceptance criteria

• Warnings publish through the LSP with warning severity and never block build/codegen by themselves.
• Exact unused local/parameter cases receive deterministic diagnostics.
• Read/write/move/drop facts prevent false “unused” classifications.
• An unsafe-to-remove binding receives a warning without an unsafe quick fix.
• A proven-safe fix is represented through the compiler source-action/change-plan seam from Add scalable auto-import code actions to the Silk LSP #178.
• Tests cover effectful initializers, affine values, writes without reads, parameters, pattern bindings, shadowing, and public declarations.
• Lint configuration and intentional suppression are specified before implementation is considered complete.

Follow-up integration

Function-contract and ordinary/effect policy warnings should use this lint framework, but their semantic analyses remain separate issues.