Tail-recursion optimisation (TCO)

[01mf02]

This PR implements tail-recursion optimisation, as announced in #111.

This change seems to massively improve the performance of paths and repeat, because both are now implemented using tail-recursion. (This can be seen by the impact on tree-paths and ex-implode.) Unfortunately, flatten seems to have become a bit slower. This is probably due to the lower performance of recurse, which is now implemented by definition instead of as native filter. Most other benchmarks remain roughly the same. ack is a bit slower, which might be because it is identified as mixed-tail-recursive (containing tail-recursive and non-tail-recursive calls).

Benchmark	n	jaq-1.1	jaq-1.2
empty	512	650	700
bf-fib	13	390	410
reverse	1048576	40	40
sort	1048576	120	120
group-by	1048576	400	380
min-max	1048576	200	180
add	1048576	450	470
kv	131072	170	160
kv-update	131072	190	180
kv-entries	131072	590	580
ex-implode	1048576	690	510
reduce	1048576	740	730
try-catch	1048576	160	160
tree-flatten	17	590	650
tree-update	17	400	400
tree-paths	17	1280	470
to-fromjson	65536	30	40
ack	7	540	570

Furthermore, this PR now enables more lazy evaluation of arrays. Previously, running

def f: 1, [f]; limit(1; f)

would result in a stack overflow, because it evaluated [f] eagerly. Now, this yields just 1, as expected.