Avoid racing between session close + adding new streams

[euank]

Until now, there was a subtle race between session.die and a new stream being created.

In session.die, we used streamMap.Each to close all streams. We also set local.goneAway in die, which in theory should mean new streams aren't added.

However, setting goneAway, iterating with Each, and adding a stream aren't adequately synchronized, so it's possible to end up in a state where we've got a goroutine already in handleSyn, about to add a new stream, and then we context switch to 'die' and cleanup existing streams, and then switch back and add the stream.

This puts the stream in a state where it will never get closed since die won't run again (due to dieOnce), however we won't process the stream properly since die closed the s.dead channel and cleaned up stuff related to that.

So, what's the fix?

Instead of using streamMap.Each during die, we use streamMap.Drain() to atomically clear the map and mark it drained, and then in Set we return whether we just added to a drained map (the race condition), and if we did we knock over the stream we just created.

The included test failed before this change.

The test was written by claude, the fix was written by me.

[bmpngrok]

Thanks for the fix! Called out one nit and one question, but I genuinely don't have a better suggestion after looking at the code that isn't a lot heavier than what you've got here

[bmpngrok]

I accept this is the best solution, but since it does open us to NPEs, was there any alternative other than the nil set? (I strongly guess not, but just double-checking. I see the several checks you had to add for m.table == nil, and that makes me nervous.)

[euank]

I mean, we could add a bool like m.frozen or such, but I feel like this struct is small and self-contained enough that it's easiest to just be careful about nils.