[andreacappelletti97]

Fixes #116.

Expressions reading module.<name>.<output> never resolved. Nothing put module in the evaluation scope, and output blocks were never evaluated. So:

• a count or for_each built from a module output dropped the whole block;
• an attribute set from one was priced at the catalog default.

Change

• Module calls are resolved before any resource is emitted. For each call, its instances are built, their own calls are resolved recursively, and their output blocks are evaluated in the module's scope.
• The outputs go into the caller's scope as module.<name>: a tuple under count, and an object keyed by each.key under for_each. The caller's locals are then resolved again before the next call.
• Calls are ordered so that one reading module.X (directly or through a local) comes after X. A cycle is broken in source order, and the parse still finishes.
• Resources are emitted afterwards from the resolved tree, in the same order as before.
• The parse budget still counts each module instance once.
• Output blocks in .tf.json files are now translated (only value).

Verification

• The repro from Module outputs aren't evaluated: for_each over module.<name>.<output> drops the block #116 through the CLI: v0.3.20 gives $41.90 with the queues dropped and both DB classes at the catalog default. This branch gives $284.26, with db.m7g.large for both and q["a"] and q["b"] present.
• New tests in module_outputs_test.go: the Module outputs aren't evaluated: for_each over module.<name>.<output> drops the block #116 repro, outputs through locals with the consumer declared first, count/for_each/nested outputs, JSON outputs driving count, and a cycle. All of them fail on main except the cycle test.
• go test ./... and golangci-lint v2.15.0 (Go 1.27.3) pass.
• BenchmarkParseLargeMonorepo is about 5% slower (one extra locals pass per module).