perf: cache and optimize struct tag processing

[tcard]

Currently, mapstructure is often slower than just JSON-marshaling and unmarshaling again.

Adding some caches improves performance significantly.

goos: darwin
goarch: arm64
pkg: github.com/go-viper/mapstructure/v2
cpu: Apple M4
                              │ /tmp/msbench/baseline.txt │        /tmp/msbench/final.txt        │
                              │          sec/op           │    sec/op     vs base                │
_Decode-10                                   1.862µ ±  0%   1.637µ ±  5%  -12.08% (p=0.001 n=10)
_DecodeBasic-10                              7.804µ ±  4%   6.071µ ±  1%  -22.21% (p=0.000 n=10)
_DecodeEmbedded-10                           6.131µ ±  6%   3.721µ ±  2%  -39.31% (p=0.000 n=10)
_DecodeMapOfStruct-10                       11.847µ ±  1%   7.250µ ± 15%  -38.80% (p=0.002 n=10)
_DecodeSliceOfStruct-10                     11.593µ ±  9%   6.742µ ± 20%  -41.85% (p=0.000 n=10)
_DecodeMetadata-10                           1.569µ ± 12%   1.228µ ±  2%  -21.73% (p=0.000 n=10)
_DecodeMetadataEmbedded-10                   7.599µ ± 10%   5.769µ ±  3%  -24.09% (p=0.000 n=10)
_DecodeTagged-10                             817.2n ±  5%   582.4n ±  5%  -28.74% (p=0.000 n=10)
_DecodeWithRemainingFields-10                1.957µ ±  3%   1.841µ ±  1%   -5.95% (p=0.000 n=10)
geomean                                      3.949µ         2.882µ        -27.01%