Correct tail recursion & improve compiler complexity

[01mf02]

In #216, @wader found an expression t that would not yield the same result as t | ..
This boiled down to a bug in tail-call optimisation (TCO).

This PR corrects this bug. At the same time, it makes compilation more efficient, in particular when there are many definitions, variables, or labels. For example:

$ (echo 'def a: 1;'; for i in `seq 100000`; do echo 'def b: a;'; done; echo b) > bla.jq
$ jaq -n -f bla.jq
1

Previously, this would have taken quadratic time (with increasing values of seq ...); now, this takes n * log n time.
On the downside, the startup time becomes slightly longer for very small programs.

$ ./bench.sh target/release/jaq-tco-*
{"name": "empty", "n": 512, "time": {"target/release/jaq-tco-bug": [0.28], "target/release/jaq-tco-fix": [0.32]}}
{"name": "bf-fib", "n": 13, "time": {"target/release/jaq-tco-bug": [0.51], "target/release/jaq-tco-fix": [0.52]}}
{"name": "defs", "n": 100000, "time": {"target/release/jaq-tco-bug": [0.04], "target/release/jaq-tco-fix": [0.05]}}
{"name": "reduce-update", "n": 16384, "time": {"target/release/jaq-tco-bug": [0.01, 0.01, 0.01], "target/release/jaq-tco-fix": [0.01, 0.01, 0.01]}}
{"name": "reverse", "n": 1048576, "time": {"target/release/jaq-tco-bug": [0.09, 0.08, 0.07], "target/release/jaq-tco-fix": [0.08, 0.08, 0.08]}}
{"name": "sort", "n": 1048576, "time": {"target/release/jaq-tco-bug": [0.16, 0.16, 0.16], "target/release/jaq-tco-fix": [0.16, 0.16, 0.15]}}
{"name": "group-by", "n": 1048576, "time": {"target/release/jaq-tco-bug": [0.55, 0.55, 0.53], "target/release/jaq-tco-fix": [0.54, 0.54, 0.54]}}
{"name": "min-max", "n": 1048576, "time": {"target/release/jaq-tco-bug": [0.25, 0.25, 0.25], "target/release/jaq-tco-fix": [0.25, 0.25, 0.26]}}
{"name": "add", "n": 1048576, "time": {"target/release/jaq-tco-bug": [0.55, 0.52, 0.54], "target/release/jaq-tco-fix": [0.53, 0.54, 0.53]}}
{"name": "kv", "n": 131072, "time": {"target/release/jaq-tco-bug": [0.13, 0.13, 0.14], "target/release/jaq-tco-fix": [0.14, 0.14, 0.14]}}
{"name": "kv-update", "n": 131072, "time": {"target/release/jaq-tco-bug": [0.15, 0.15, 0.15], "target/release/jaq-tco-fix": [0.15, 0.15, 0.14]}}
{"name": "kv-entries", "n": 131072, "time": {"target/release/jaq-tco-bug": [0.62, 0.61, 0.61], "target/release/jaq-tco-fix": [0.61, 0.60, 0.61]}}
{"name": "ex-implode", "n": 1048576, "time": {"target/release/jaq-tco-bug": [0.56, 0.56, 0.56], "target/release/jaq-tco-fix": [0.56, 0.53, 0.54]}}
{"name": "reduce", "n": 1048576, "time": {"target/release/jaq-tco-bug": [0.82, 0.80, 0.83], "target/release/jaq-tco-fix": [0.81, 0.81, 0.81]}}
{"name": "try-catch", "n": 1048576, "time": {"target/release/jaq-tco-bug": [0.35, 0.35, 0.36], "target/release/jaq-tco-fix": [0.35, 0.36, 0.36]}}
{"name": "tree-contains", "n": 23, "time": {"target/release/jaq-tco-bug": [0.07, 0.07, 0.07], "target/release/jaq-tco-fix": [0.07, 0.07, 0.07]}}
{"name": "tree-flatten", "n": 17, "time": {"target/release/jaq-tco-bug": [0.89, 0.89, 0.90], "target/release/jaq-tco-fix": [0.89, 0.88, 0.86]}}
{"name": "tree-update", "n": 17, "time": {"target/release/jaq-tco-bug": [0.72, 0.72, 0.71], "target/release/jaq-tco-fix": [0.72, 0.72, 0.72]}}
{"name": "tree-paths", "n": 17, "time": {"target/release/jaq-tco-bug": [0.46, 0.47, 0.45], "target/release/jaq-tco-fix": [0.47, 0.46, 0.47]}}
{"name": "to-fromjson", "n": 65536, "time": {"target/release/jaq-tco-bug": [0.04, 0.04, 0.04], "target/release/jaq-tco-fix": [0.04, 0.04, 0.04]}}
{"name": "ack", "n": 7, "time": {"target/release/jaq-tco-bug": [0.53, 0.53, 0.53], "target/release/jaq-tco-fix": [0.49, 0.49, 0.50]}}
{"name": "range-prop", "n": 128, "time": {"target/release/jaq-tco-bug": [0.38, 0.38, 0.38], "target/release/jaq-tco-fix": [0.38, 0.39, 0.39]}}
{"name": "cumsum", "n": 1048576, "time": {"target/release/jaq-tco-bug": [0.31, 0.33, 0.32], "target/release/jaq-tco-fix": [0.33, 0.34, 0.32]}}
{"name": "cumsum-xy", "n": 1048576, "time": {"target/release/jaq-tco-bug": [0.54, 0.54, 0.53], "target/release/jaq-tco-fix": [0.55, 0.54, 0.53]}}

Here, we can see slightly worse values for empty, but improvements for ack.