[tailscale] runtime: add TailscaleReadStackStats for per-goroutine stack info

[bradfitz]

We're working on reducing the standing memory of servers holding
millions of idle long-polling connections. A big part of that is
knowing which stack size class those goroutines' stacks are in, how
much of the stack they actually use while parked, and whether the
stacks grew during connection setup and were later shrunk back by the
GC. The runtime doesn't expose any of that.

Add runtime.TailscaleReadStackStats, modeled on ReadMemStats, which
fills in a TailscaleStackStats for the calling goroutine: the current
stack size (its size class), the bytes in use at the point of the
call, the largest size the stack has ever been, and saturating counts
of growths and shrinks.

The runtime doesn't track a true high-water mark of stack usage, and
doing so would cost work on every call, so MaxSize (the peak allocated
size) is the proxy: a stack only doubles when the previous size
overflowed, so peak usage was at least roughly MaxSize/2. Comparing it
against Size shows whether the GC has already shrunk a goroutine's
stack back down.

The three backing fields on g are uint8s placed in what was alignment
padding before g.sig, so g does not grow. They're reset in newproc1
because gs are recycled through the free list. Growths are recorded
in newstack and shrinks in shrinkstack. The call itself takes no locks
and costs about 7ns.

Updates #184

[tomhjp]

Nice tests/benchmark! LGTM