Fix OCaml compiler regressions

[miselin]

Summary

• restore true short-circuit evaluation and left associativity for multiplicative operators
• allow composable pointer/type postfixes and mutation through casted pointer expressions
• preserve initializer-less public globals as external declarations
• emit correctly typed constant global initializers and keep startup-initialized data writable in LLVM
• report duplicate function definitions before malformed LLVM reaches code generation
• strengthen parser, LLVM IR, semantic, and runtime regression coverage

Why

I attempted a port of DOOM from C to Haven, and discovered several rough edges and pure compiler bugs along the way. Some of these were regressions from the OCaml branch, but some were carried over from the previous C implementation of the compiler.

This also revealed some limits in the test coverage. For example, the previous short-circuit test used division by zero as an unobservable RHS computation, which allowed eager evaluation to escape detection. That test has been updated to explicitly flush out incorrect short-circuit behavior and avoid regressing again.

Validation

• dune runtest
• dune build @all
• runtime regression cases at -Os, -O0, -O1, -O2, and -O3
• git diff --check