Optimize analysis backend

[Techatrix]

based on top of #1057
fixes #536

I've benchmarked the total time required to generate semantic tokens for Sema.zig 20 times in a row.

Command	Mean [s]	Min [s]	Max [s]	Relative	per request	speedup 🚀
./bench-new-debug	8.273 ± 0.012	8.257	8.291	6.07 ± 0.08	413.65ms	16.7x
./bench-new-fast	1.364 ± 0.017	1.346	1.383	1.00	68.2ms	21.6x
./bench-new-safe	1.554 ± 0.012	1.537	1.569	1.14 ± 0.02	77.7ms	21.1x
./bench-new-small	1.964 ± 0.008	1.954	1.975	1.44 ± 0.02	98.2ms	17.8x
./bench-old-debug	138.830 ± 0.898	137.611	139.700	101.81 ± 1.43	6941.5ms
./bench-old-fast	29.566 ± 0.937	28.677	30.757	21.68 ± 0.74	1478.3ms
./bench-old-safe	32.906 ± 1.089	31.477	34.024	24.13 ± 0.85	1645.3ms
./bench-old-small	35.037 ± 0.845	34.158	35.969	25.69 ± 0.70	1751.8ms

The actual changes that improved performance are relatively small:

• Calls to resolveTypeOfNode are now cached so they never have to be recomputed
• scopes in DocumentScope now stores their parent and child scopes so that variable lookup becomes faster