doh! fixes issue 92 again.

tentative commit. this should solve issue 92, but isn't working, yet.

the idea is to taint any memoized follow set that hit a recursive loop,
directly or indirectly. the recursive calls (h_follow_rec) are allowed to use
these intermediates, but propagate the taint. the top-level routine (h_follow)
places the original taint on its result and may remove it at the end. it will
not return memoized results that are tainted (but compute them afresh).

without it, our test for issue 92 is no good (will always pass).
we could probably officially put it back if we simplify "dead"
symbols (that have no productions) out of the grammar beforehand.

there is a case where fs can be empty: when reducing by lhs would lead to
certain parse failure, by means of h_nothing_p() for instance.

fixes issue 83.

cf.
hammer/hammer#83

this makes the computation more wasteful but avoids issue 92. intermediate
results could end up in the memoization table and be treated as final by
later calls to h_follow(). this problem could appear or not depending on the
order of nonterminals (i.e. pointers) in a hashtable.

hammer/hammer#92

development

update started from.

fixes issue 91

"assertion failure with ambiguous LALR compiles"
hammer/hammer#91

i was trying to find out why the test suite suddenly insisted on being so
noisy. turns out glib now does TAP by default and all messages get printed,
--verbose or not. the old mode is considered deprecated.

cf.
https://en.wikipedia.org/wiki/Test_Anything_Protocol
https://gitlab.gnome.org/GNOME/glib/commit/ad175766575b23be920f41cdb1c350447c743347
https://gitlab.gnome.org/GNOME/glib/commit/19aafc4ca4dba2b63d24f01d23890ff30ef9b133

this omission has nagged me a number of times and i think it must have been a
simple oversight.

this way the assert will fire from the user's code rather than the one in
glue.c:h_seq_index() which is not informative without a stack trace.

also includes commented-out tests for functionality that i discovered we never
provided, namely position and length information at the granularity of
HParsedTokens. at the moment, these fields always seem to be 0. we should
probably fix that or get rid of them. in the latter case it might make sense
to move this information to HParseResult, so the user can get access to it via
h_action() as needed.

the regex backend already returned 2 for an internal failure, contrary to
documentation, so this mainly brings the latter in line with existing behavior.

at the same time, fix a bug: h_compile would previously return "false" (0) on
exceptions (longjmp).

make use of our new freedom to signal different error conditions from the
context-free backends, notably "incompatible combinator" vs. "unresolvable
conflict".

Provide an alternate interface to the arena allocator to avoid wasting time on unnecessary calls to memset()

Eliminate unnecessary hash computations in the packrat backend (roughly 2.5x speed in one benchmark)

A working h_drop_from which rewrites its underlying sequence as a sequence containing some ignored parsers.