Your list comprehension is in essence just syntactical sugar for:
import Control.Monad(guard)
boomBangs :: Integral i => [i] -> [String]
boomBangs xs = do
    x <- xs
    guard (odd x)
    return (if x < 10 then "BOOM!" else "BANG!")
This is thus a do expression [Haskell report], and as the report says, it is syntactical sugar. It is syntactical sugar for:
boomBangs xs = xs >>= \x -> (guard (odd x) >> return (if x < 10 then "BOOM!" else "BANG!"))
For a list the Monad instance is defined as:
instance Monad [] where
    (>>=) = flip concatMap
    return x = [x]
Furthermore the guard is defined as:
guard :: Monad m => Bool -> m ()
guard True = pure ()
guard False = empty
and the default implementation of (>>) is:
(>>) :: Monad m => m a -> m b -> m b
(>>) u v = u >>= \_ -> v
So the boomBangs is basically implemented as:
boomBangs xs = concatMap (\x -> (guard (odd x) >>= \_ -> [if x < 10 then "BOOM!" else "BANG!"])) xs
             = concatMap (\x -> concatMap (\_ -> [if x < 10 then "BOOM!" else "BANG!"]) guard (odd x)) xs
Since for a list, the guard can be specialized to:
-- guard for the Monad []
guard :: Bool -> [()]
guard True = [()]
guard False = []
It thus means that if the guard gets a True, it returns a singleton list, and for False an empty list. This thus means that given the guard holds, the concatMap (\_ -> [if x < 10 then "BOOM!" else "BANG!"]) will return the content in the [if x < 10 then "BOOM!" else "BANG!"], if the guard fails, it will return an empty list. The guard acts thus as some sort of filter.
So now what is the x <-. If we look at how the do-expressions are desugared, an x <- foo, corresponds to the foo >>= \x -> ....
For list comprehensions, x <- ... acts as some sort of "enumerator": it will enumerate over all elements in the list, and x will each time obtain one of the elements in the list to do further processing.