Merge Basic Blocks if possible

[reiner-dolp]

Currently, through optimization, sometimes nodes are split into a suboptimal amount of basic blocks, which might reduce the efficiency of subsequent optimizations.

---

The following is an example:

---
optimizations: [ConstantFolding,UnreachableCodeElimination]
expect:
    identical-to:
        is: |
          class UCE_If {
            public static void main(String[] args) {
              System.out.println(1);
            }
          }
        
stdout:
    is: "1\n"
---
class UCE_If {
  public static void main(String[] args) {
    if (true) {
        System.out.println(1);
    } else {
        System.out.println(0);
    }
  }
}

The ASM for the hand optimized routine is:

# -- Begin  mj_main
	.p2align 4,,15
	.globl mj_main
	.type	mj_main, @function
mj_main:
/*.L0:*/                          /* Block BB[108:2] preds: none, freq: 1.000 */
	pushq %rbp                       /* amd64_push_reg T[273:74] */
	mov %rsp, %rbp                   /* be_Copy Lu[276:77] */
	movl $0x1, %edi                  /* amd64_mov_imm Lu[207:11] */
	call mjrt_system_out_println@PLT    /* amd64_call T[208:12] */
	leave                            /* amd64_leave T[269:70] */
	ret                              /* amd64_ret X[216:20] */
	.size	mj_main, .-mj_main
# -- End  mj_main

The compiler optimized routine has unnecessary basic blocks, resulting in fallthroughs to be visible in the ASM:

mj_main:
/*.L800:*/                        /* Block BB[176:2] preds: none, freq: 1.000 */
    pushq %rbp                       /* amd64_push_reg T[635:78] */
    mov %rsp, %rbp                   /* be_Copy Lu[638:81] */
    /* fallthrough to .L801 */       /* amd64_jmp X[578:24] */
/*.L801:*/                        /* Block BB[189:10] preds: .L800, freq: 1.000 */
    movl $0x1, %edi                  /* amd64_mov_imm Lu[567:13] */
    call mjrt_system_out_println@PLT    /* amd64_call T[568:14] */
    /* fallthrough to .L802 */       /* amd64_jmp X[577:23] */
/*.L802:*/                        /* Block BB[215:9] preds: .L801, freq: 1.000 */
    leave                            /* amd64_leave T[631:74] */
    ret                              /* amd64_ret X[576:22] */
    .size    mj_main, .-mj_main
# -- End  mj_main

[reiner-dolp]

We also emit empty blocks during firm graph construction:

comprakt/compiler-lib/src/firm/method_body_generator.rs

Line 753 in 9fdd9d5

// FIXME: we're constructing an empty block here because we are only

[problame]

This would actually be "useful" for the backend, as some of our more esoteric crashes stem from unnecessarily split basic blocks (e.g. the problems in #217 with Cmp and Cond being spread over 2 BBs even though they could be in one).
Technically, it shouldn't matter for the backend, though.