This is a big commit, but it deletes lots of code and simplifies a lot.
* Resolving the template funcs at execution time means we don't have to create template clones per site
* Having a custom map resolver means that we can remove the AST lower case transformation for the special lower case Params map
Not only is the above easier to reason about, it's also faster, especially if you have more than one language, as in the benchmark below:
```
name old time/op new time/op delta
SiteNew/Deep_content_tree-16 53.7ms ± 0% 48.1ms ± 2% -10.38% (p=0.029 n=4+4)
name old alloc/op new alloc/op delta
SiteNew/Deep_content_tree-16 41.0MB ± 0% 36.8MB ± 0% -10.26% (p=0.029 n=4+4)
name old allocs/op new allocs/op delta
SiteNew/Deep_content_tree-16 481k ± 0% 410k ± 0% -14.66% (p=0.029 n=4+4)
```
This should be even better if you also have lots of templates.
Closes#6594
When run under gccgo, the test looks for the name that gccgo gives to
a thunk method. This name is not normally visible, but can be seen
when using reflect.FuncForPC as this code does. That name changed in
https://golang.org/cl/89555. Change the test to work with both the
old name "$thunk0" and the new name "thunk0".
Fixesgolang/go#28669
As a first step to remove the hard ties between `tplimpl` and the different namespace packages.
The `lang` package is used as the first example use case.
See #3042
