types: alpha-rename imported schemes into the consumer's id space

[simontreanor]

Fixes #26.

Imported schemes were seeded into the consumer''s env carrying the exporting module''s raw variable ids, while every module allocates its own ids from the same counter (RESERVED_VARS). instantiate survives that on its own — it refreshes the quantified vars before substituting — but env_free_vars applies the local substitution to every env scheme, so a single id collision silently rewrote the imported scheme and leaked local variables into the "free in the environment" set. generalize then refused to quantify them.

The consequences compound:

• the importing module''s own let f a b = (a, b) goes monomorphic, so a second use at another type is rejected;
• that binding exports un-quantified, so consumers one module further along share a single type variable across every use site.

Both were observed in a real project: an exported refill came out as (List ''a) -> (List ''a) -> (List ''a, List ''a) with nothing quantified, and the imported scheme whose collision caused it had been rewritten from int -> List a -> List a into int -> List ((List ''a) -> List ''a) -> List ((List ''a) -> List ''a).

Fix

Infer::refresh_scheme alpha-renames an imported scheme into the consumer''s variable space — type, unit, num and effect ids, carrying ord_vars across — applied to imported values and constructors as they are seeded, in sorted order so the ids do not depend on HashMap iteration. Free ids are renamed as well as quantified ones: a scheme should arrive closed, but a free id is exactly the case that must not alias a local variable.

Tests

Two regression tests in tests/project.rs, both verified to fail on main:

• an_importing_module_keeps_its_own_functions_polymorphic — the direct case; Main does not even use the import.
• an_under_generalized_export_does_not_cascade_to_a_consumer — the contagious case, three modules deep.

cargo test green (all suites), cargo clippy --all-targets clean, cargo fmt applied. INTERNALS.md updated: the old soundness note claimed closed schemes were sufficient, which is where the reasoning went wrong.