Allow one layer of unbalanced if to be inlined

[limepoutine]

This now gives small performance benefits with std.random unittests (dmd -inline -main -O -unittest -version=StdUnittest random.d).

Benchmark 1: ./random_master
  Time (mean ± σ):     451.5 ms ±   5.1 ms    [User: 450.3 ms, System: 3.1 ms]
  Range (min … max):   443.7 ms … 465.3 ms    20 runs

Benchmark 2: ./random_pr
  Time (mean ± σ):     444.1 ms ±   4.1 ms    [User: 442.9 ms, System: 3.1 ms]
  Range (min … max):   438.6 ms … 455.5 ms    20 runs
 
Summary
  ./random_pr ran
    1.02 ± 0.01 times faster than ./random_master

[github-actions]

DMD perf check

Metric	Base	PR	Δ
dmd binary size (stripped)	6.87 MB	6.88 MB	+0.11%

All measurements

Metric	Base	PR	Δ
compile hello.d (instr)	213.4 M	213.4 M	+0.001%
compile hello.d -O -release (instr)	231.7 M	231.7 M	0.000%
compile Phobos (instr)	5,124.4 M	5,124.0 M	-0.007%
compile Phobos codegen (instr)	1,473.6 M	1,473.1 M	-0.036%
compile vibe.d (instr)	15,116.1 M	15,118.0 M	+0.012%
dmd binary size (stripped)	6.87 MB	6.88 MB	+0.11%
hello binary size (stripped)	0.72 MB	0.72 MB	0.00%
peak RSS (compile hello.d)	43.14 MB	43.14 MB	-0.02%
peak RSS (compile Phobos)	619.0 MB	617.7 MB	-0.20%
peak RSS (compile vibe.d)	1918 MB	1917 MB	-0.02%
compile dmd itself (wall)	12.1 s	12.1 s	+0.31%
compile hello.d (wall)	64.1 ms	63.7 ms	-0.68%
compile Phobos (wall)	1,580 ms	1,589 ms	+0.55%

_{033916e vs merge-base 9af84e9 · about these metrics}