Use lld instead of avr-gcc as the linker

This fixes the [build error] caused by the `avr-gcc` (used as linker)
not being available in the Rust CI. This is a viable solution, which
shows the wrong/right behavior and, since no functions from `libgcc` are
called, does not produce errors. This was discussed [here]. Another
small problem is, that `lld` doesn't link the correct startup-code by
default. This is not a problem for this test (since it does not actually
use anything the startup code is needed for (no variables, no stack, no
interrupts)), but this causes the `main`-function to be removed by the
default flag `--gc-sections`. Therefore the `rmake`-driver also adds the
linker flag `--entry=main` to mark the `main`-function as the entry
point and thus preventing it from getting removed. The code would work
on a real AVR device.

[build error]: rust-lang#131755 (comment)
[here]: rust-lang#131755 (comment)

tests/run-make/avr-rjmp-offset/rmake.rs

@@ -17,6 +17,13 @@ fn main() {
         .opt_level("s")
         .panic("abort")
         .target("avr-unknown-gnu-atmega328")
+        // normally one links with `avr-gcc`, but this is not available in CI,
+        // hence this test diverges from the default behavior to enable linking
+        // at all, which is necessary for the test (to resolve the labels). To
+        // not depend on a special linker script, the main-function is marked as
+        // the entry function, causing the linker to not remove it.
+        .linker("lld")
+        .link_arg("--entry=main")
         .output("compiled")
         .run();
 
@@ -35,6 +42,7 @@ fn main() {
     // fore the relative jump has an impact on the label offset. Old versions
     // of the Rust compiler did produce a label `rjmp .-4` (misses the first
     // instruction in the loop).
+    assert!(disassembly.contains("<main>"), "no main function in output");
     disassembly
         .trim()
         .lines()