fix: Ensure submodules compile before programs for LFortran direct mode (#1252)

[ijatinydv]

Resolves #1252

Currently, FPM's topological sort allows programs and submodules sharing a parent module to be scheduled in the same OpenMP region. For compilers using direct compilation (like LFortran), this creates a race condition where the main program attempts to compile before all .smod files are generated.

The Fix
Added a post-resolution pass in resolve_module_dependencies (fpm_targets.f90) that establishes a blanket dependency edge from all FPM_UNIT_PROGRAM and FPM_UNIT_TEST targets to all FPM_UNIT_SUBMODULE targets.

Why a blanket rule instead of parent-matching?
Parent-matching fails on transitive dependencies (e.g., main -> mod_A -> mod_B -> submod_B misses the indirect submodule because modules_used only contains mod_A). Since programs and tests are terminal nodes in the build graph, adding blanket edges has zero performance cost and guarantees correct compilation order.

Tests Added
Added 5 new tests to test_module_dependencies.f90 to verify:

1. Basic program/submodule ordering.
2. Transitive dependency resolution.
3. App-scope submodules.
4. Multi-program scenarios.
5. Tree-shaking: Verified that the new blanket rule does not prevent unused submodules from being pruned.

cc @certik