... preparing to add a nextTreeId as well.

This removes some technical debt: nextId was a global counter, so
unreliable under parallelism. That was OK so far since it was only
intended for debugging. But now we want to give more significance to
tree ids, since they will be used to determine the source file of
a tree. Hence, we need to put this aspect in order by making nextTreeId
a context field. This entailed a lot of changes, so no wonder it was
not done before.

It's very important that tree nodes don't capture contexts. The @transientParam
annotation makes sure they don't.

Tree ids are now allocated in chunks. All ids in the same chunk
were produced by contexts with the same source file.

Inlined sources might not come from a compilation units that's currently compiled.

Copies should have the same source attribute as originals.

Serialize source file changes in positions section

Tree ids survive compilation runs, need to be allocated globally.

Pickle SOURCE position for roots, so that unpickler get the'right source when
constructing trees.

Generate a position entry if sourcefiles change.

Used to depend on file's path, but that does not work for VirtualFiles
that might be different even though they have the same path.

Use interned strings for the comparison of their paths.

Reduce concurrent hashmap lookups in TreeIds by caching the last result in SourceInfo.
This reduces lookups for typer/*.scala from 500K+ to 7.6K.

Deemphasize AbstractFile, drop TreeIds.

When set, prints full source positions, not just file + line.

The decision when to pickle a position is a time sink. Try to optimize
this by not going through Positioned#{initialSpan,elemSource} functionality.

Single traversal to set id and span.

It's called more often now that we generation source positions. Time
for some caching.

No need to do the wrapping