Decode primitive types directly, without intermediate INode

Fix TSize (de)coding on a way there.

This improves performance from:

```
BenchmarkDecode<TFlatStructLite>       24437 ns        24433 ns        29036 items_per_second=40.9277k/s
BenchmarkDecode<TFlatStructHeavy>      24807 ns        24804 ns        29581 items_per_second=40.3154k/s
BenchmarkDecode<TDeepStructLite>       88384 ns        88365 ns         7578 items_per_second=11.3167k/s
BenchmarkDecode<TDeepStructHeavy>      42440 ns        42432 ns        16215 items_per_second=23.5672k/s
```

To:

```
BenchmarkDecode<TFlatStructLite>       10463 ns        10461 ns        67494 items_per_second=95.5953k/s
BenchmarkDecode<TFlatStructHeavy>      10816 ns        10814 ns        66032 items_per_second=92.4694k/s
BenchmarkDecode<TDeepStructLite>       72614 ns        72602 ns         9776 items_per_second=13.7737k/s
BenchmarkDecode<TDeepStructHeavy>      26924 ns        26919 ns        26204 items_per_second=37.1486k/s
```

Part 1 of [https://nda.ya.ru/t/E9Ji-1bm7Gwv6c](https://nda.ya.ru/t/3w3MLYZB7Gwv6d
commit_hash:de01ea3b3e667f0e090bf3c61df79726917a6ebd

0 files changed, 0 insertions, 0 deletions