Run all linters per package. It allows unloading package data when it's
processed. It dramatically reduces memory (and CPU because of GC) usage.
Relates: #337
Use go/packages instead of x/tools/loader: it allows to work
with go modules and speedups loading of packages with the help
of build cache.
A lot of linters became "fast": they are enabled by --fast now and
work in 1-2 seconds. Only unparam, interfacer and megacheck
are "slow" linters now.
Average project is analyzed 20-40% faster than before if all linters are
enabled! If we enable all linters except unparam, interfacer and
megacheck analysis is 10-20x faster!
Use build.Import instead of manual parser.ParseFile and paths traversal. It allows:
1. support build tags for all linters.
2. analyze files only for current GOOS/GOARCH: less false-positives.
3. analyze xtest packages (*_test) by golint: upstream golint and
gometalinter can't do it! And don't break analysis on the first xtest
package like it was before.
4. proper handling of xtest packages for linters like goconst where
package boundary is important: less false-positives is expected.
Also:
1. reuse AST parsing for golint and goconst: minor speedup.
2. allow to specify path (not only name) regexp for --skip-files and
--skip-dirs
3. add more default exclude filters for golint about commits:
`(comment on exported (method|function)|should have( a package)?
comment|comment should be of the form)`
4. print skipped dir in verbose (-v) mode
5. refactor per-linter tests: declare arguments in comments, run only
one linter and in combination with slow linter
