`do_collect` needs optimized

[Others]

We can probably allocate less in this method, and rayon could be used to speed up the scanning process.

[Others]

We're using rayon now. The biggest future optimization is this one for determining what's rooted:

• Track an "invalidation number" for each handle, a u64 that's incremented whenever a lockout is taken
• For each object, take lockout if possible, then mark each interior handle as not rooted, recording invalidation num
• Then go through handles. Mark as non-rooted if marked as non-rooted and invalidation number not incremented (this is necessarily conservative. If it is marked non-rooted and the invalidation number was not incremented, it’s still effectively in place—only the set of roots can increase)
• Then freeze everything so scanning can see a consistent view of the system
• Then proceed as usual