database/sql: fix race when canceling queries immediately

Previously the following could happen, though in practice it would
be rare.

Goroutine 1:
	(*Tx).QueryContext begins a query, passing in userContext

Goroutine 2:
	(*Tx).awaitDone starts to wait on the context derived from the passed in context

Goroutine 1:
	(*Tx).grabConn returns a valid (*driverConn)
	The (*driverConn) passes to (*DB).queryConn

Goroutine 3:
	userContext is canceled

Goroutine 2:
	(*Tx).awaitDone unblocks and calls (*Tx).rollback
	(*driverConn).finalClose obtains dc.Mutex
	(*driverConn).finalClose sets dc.ci = nil

Goroutine 1:
	(*DB).queryConn obtains dc.Mutex in withLock
	ctxDriverPrepare accepts dc.ci which is now nil
	ctxCriverPrepare panics on the nil ci

The fix for this is to guard the Tx methods with a RWLock
holding it exclusivly when closing the Tx and holding a read lock
when executing a query.

Fixes #18719

Change-Id: I37aa02c37083c9793dabd28f7f934a1c5cbc05ea
Reviewed-on: https://go-review.googlesource.com/35550


Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>