precompile: compile inlinable methods that were inferred

The code generation policy here previously assumes that inlinable code
almost never ends up being `invoke`d, but it has the unfortunate side
effect that making more code inline-eligible means fewer entry points
to your code end up compiled.

Intuitively I think users expect function calls that ran during
pre-compilation to be compiled and available (regardless of the context
they are called from), which requires us to pre-compile these

This change may not be viable due to the increase in code size, but I
wanted to open it to see what the effects are.

src/precompile_utils.c

@@ -187,9 +187,8 @@ static int precompile_enq_specialization_(jl_method_instance_t *mi, void *closur
         }
         else if (jl_atomic_load_relaxed(&codeinst->invoke) != jl_fptr_const_return) {
             jl_value_t *inferred = jl_atomic_load_relaxed(&codeinst->inferred);
-            if (inferred &&
-                inferred != jl_nothing &&
-                (jl_options.compile_enabled != JL_OPTIONS_COMPILE_ALL && jl_ir_inlining_cost(inferred) == UINT16_MAX)) {
+            if (inferred && inferred != jl_nothing && jl_options.compile_enabled != JL_OPTIONS_COMPILE_ALL &&
+                (jl_options.trim == JL_TRIM_NO || jl_ir_inlining_cost(inferred) == UINT16_MAX)) {
                 do_compile = 1;
             }
             else if (jl_atomic_load_relaxed(&codeinst->invoke) != NULL || jl_atomic_load_relaxed(&codeinst->precompile)) {