The computation of nullable is broken: it doesn't handle empty

RHS's properly.
* tests/torture.at (GNU AWK Grammar): New.
* tests/sets.at (Nullable): New.
* src/nullable.c (set_nullable): Instead of blindly looping over
`ritems', loop over the rules, and then over their rhs's.
Work around Autotest bugs.
* src/warshall.c (bitmatrix_print): Don't use `+--+' as table
frame, because Autotest understand lines starting with a `+' as
traces from the shell.  Then, they are not processed properly.
Admittedly an Autotest bug, but we don't have time to wait for
Autotest to catch up.
* tests/regression.at (Broken Closure): Adjust to the new table
frames.
Move to...
* tests/sets.at: here.

tests/regression.at

@@ -605,78 +605,3 @@ AT_CLEANUP
 
 AT_TEST_CPP_GUARD_H([input/input])
 AT_TEST_CPP_GUARD_H([9foo])
-
-
-## ---------------- ##
-## Broken Closure.  ##
-## ---------------- ##
-
-# TC was once broken during a massive `simplification' of the code.
-# It resulted in bison dumping core on the following grammar (the
-# computation of FIRSTS uses TC).  It managed to produce a pretty
-# exotic closure:
-#
-# TC: Input
-#
-#    01234567
-#   +--------+
-#  0| 1      |
-#  1|  1     |
-#  2|   1    |
-#  3|    1   |
-#  4|     1  |
-#  5|      1 |
-#  6|       1|
-#  7|        |
-#   +--------+
-#
-# TC: Output
-#
-#    01234567
-#   +--------+
-#  0| 1      |
-#  1| 111    |
-#  2| 111    |
-#  3| 1111   |
-#  4| 111 1  |
-#  5| 111  1 |
-#  6| 111   1|
-#  7| 111    |
-#   +--------+
-#
-# instead of that below.
-
-AT_SETUP([Broken Closure])
-
-AT_DATA([input.y],
-[[%%
-a: b
-b: c
-c: d
-d: e
-e: f
-f: g
-g: h
-h: 'h'
-]])
-
-AT_CHECK([bison --trace input.y 2>&1 |
-          sed -n '/^TC: Output BEGIN/,/^TC: Output END/p'],
-         [0],
-[[TC: Output BEGIN
-           @&t@
-   01234567
-  +--------+
- 0| 1111111|
- 1|  111111|
- 2|   11111|
- 3|    1111|
- 4|     111|
- 5|      11|
- 6|       1|
- 7|        |
-  +--------+
-TC: Output END
-]])
-
-AT_CLEANUP