Major performance regression compared to `text/template`

In our templates we see a major regression when using `yamltemplate`.

```
name old time/op new time/op delta
Test-48 324µs ± 2% 2360µs ± 2% +628.58% (p=0.008 n=5+5)

name old alloc/op new alloc/op delta
Test-48 52.2kB ± 0% 724.3kB ± 0% +1287.50% (p=0.008 n=5+5)

name old allocs/op new allocs/op delta
Test-48 794 ± 0% 11722 ± 0% +1376.32% (p=0.029 n=4+4)
```

Given yaml templating is a hot-path in our code, this makes the library not viable for usage as-is.

Additionally, the global locking mechanism used, while not explicitly measured, is also likely a problem for us as we do templating in many concurrent goroutines.