Optimize hot paths of encoding and decoding

[letko-dmitry]

Builds on #62: only the last 8 commits are new. Each change is a separate commit, measured on its own, so any of them can be dropped.

Encoding and decoding get about 15% faster, decoding Foundation dictionaries 19%. Behavior doesn't change.

• Small functions that run for every value are inlined, and stored arguments are consumed rather than copied (−1%).
• No more linear key scan in contains, indirect boxes and a temporary array per coding path (−12% encoding, −6% decoding).
• The keyed decoding container is final, as nothing subclasses it (−1.6% decoding).
• NSNull is checked by type, so values aren't bridged to Objective-C just to find out they aren't null (−5% decoding).
• With the default float strategy, numbers aren't cast to String first (−1…2% decoding).
• Unkeyed containers reserve room for two elements, which saves a reallocation in most of them (−2.6% encoding).
• decodeIfPresent looks a key up once instead of three times (−0.5% decoding).
• Foundation numbers, strings and data are bridged directly instead of through dynamic casts (−6% decoding Foundation dictionaries).

Measured in retired instructions with alternating A/B runs, release builds on an M4; noise is within ±0.3%.