Introduce earlydecoder & refactor SchemaMerger to integrate w/ schema cache

[radeksimko]

This is to support an upcoming LS-side change described in hashicorp/terraform-ls#354

Previously "early decoding" to find out provider requirements was actually done just prior to schema merging. This PR formalizes this stage and decouples the logic into its own package earlydecoder which can be also used independently and more efficiently by the LS.

This in turn enables us to:

• actually implement cache schema, where 1 provider is 1 entry and lookups are done individually on per-provider basis
• (later) expose requirements without further schema merging (e.g. display those requirements in the editor UI somehow)
• (later) better understand when/whether requirements have changed and e.g. whether we should nudge the user for re-init

Conversion from tfjson to hcl-lang.schema was also decoupled, so that the LS can perform this conversion early and store the schema in the desired format already in memdb internally.

[aeschright]

🚀