Improve performance of content collision detection

[erikgeiser]

This issue was raised by @mjmjelde in #385.

In ExpandContentGlobs we try building the content list by appending only elements with unique destinations. This increases the complexity in way that it does not scale for large amount of files. I hacked together a POC solution where the collision detection is performed once near the end before sorting. The benchmarks are promising:

========= 10000 Content Entries =========
goos: darwin
goarch: amd64
pkg: github.com/goreleaser/nfpm/v2/files
cpu: Intel(R) Core(TM) i5-6267U CPU @ 2.90GHz
BenchmarkExpandContentGlobs-4       	       1	8943350742 ns/op	5948665080 B/op	 2963056 allocs/op
BenchmarkExpandContentGlobsFast-4   	       2	 690795220 ns/op	503257972 B/op	 1260368 allocs/op
PASS
ok  	github.com/goreleaser/nfpm/v2/files	11.305s

========= 100000 Content Entries =========
goos: darwin
goarch: amd64
pkg: github.com/goreleaser/nfpm/v2/files
cpu: Intel(R) Core(TM) i5-6267U CPU @ 2.90GHz
BenchmarkExpandContentGlobs-4       	       1	1116330817406 ns/op	545598423120 B/op	185560019 allocs/op
BenchmarkExpandContentGlobsFast-4   	       1	25918381508 ns/op	41183230936 B/op	12604194 allocs/op
PASS
ok  	github.com/goreleaser/nfpm/v2/files	1142.497s

For large packages, this can be more than a 40x speedup.

I'll submit a PR if I find the time to clean up the very hacky POC.

[erikgeiser]

See also #393