dan: "I probably wouldn't expect those mappings to cause much
trouble at all for legitimate users at the end of the day"

shivaram: [concurs]

'm' skeletonizes to 'rn' (but is exempted by the isBoring check),
but the fullwidth 'ｍ' does not skeletonize to anything. The root cause
of this is the (still unexplained) patchiness of the skeleton mapping
for fullwidth -> standard-width Latin characters; the fix is to perform
width mapping first, before either skeletonization or isBoring.

Partially fixed earlier in #467

now that we have client.destroyed again, we don't need to test the
"0 sessions && brb disabled" condition in both AddSession() and
destroy()

Use []uint32 in bitset instead of []uint64, because it's harder to guarantee
64-bit alignment of []uint64 than I had realized:

https://go101.org/article/memory-layout.html

Fixes #519

Quit messages aren't supposed to clobber each other under any
circumstances, but actually only the default "Connection closed"
message was non-clobbering, all others would clobber.

Also ensure that people can remove their own persistent modes via CS AMODE

Fixes #505

Setting `closed` on the socket (which can mean either "the socket is broken"
or "we should close the socket at the next opportunity") was racing against
the final write. Even if socket.closed is true, we should attempt to send
buffered message data to the socket, before we send the `finalData` and
actually call `Close`.