Improve startup time (regression)

[sharkdp]

The new type checker from #443 introduced a severe regression in startup times:

hyperfine --warmup 3 --reference "./numbat-master -e '1+2'" "./numbat-1.12 -e '1+2'" --export-markdown -

Command	Mean [ms]	Min [ms]	Max [ms]	Relative
./numbat-master -e '1+2'	78.9 ± 1.2	77.1	83.8	2.28 ± 0.09
./numbat-1.12 -e '1+2'	34.5 ± 1.2	32.4	42.5	1.00

We should look into potential measures to improve this again.

If the constraint checker is the culprit, we might want to look into de-duplication of constraints before attempting to solve them. But it might be something else entirely (e.g. the mutation/traversal/cloning of ASTs during elaboration)

[sharkdp]

Thanks to the changes in #543 by @GuerricChupin, we've moved a significant step in the right direction:

Command	Mean [ms]	Min [ms]	Max [ms]	Relative
./numbat-master -e '1+2'	59.9 ± 1.2	57.9	64.3	1.78 ± 0.08
./numbat-1.12 -e '1+2'	33.7 ± 1.3	32.6	44.8	1.00

[sharkdp]

Thanks to the changes in #537 by @irevoire (which are not related to the type checker, but still), we've moved another significant step in the right direction:

Command	Mean [ms]	Min [ms]	Max [ms]	Relative
./numbat-master -e '1+2'	47.0 ± 0.9	45.5	50.3	1.37 ± 0.04
./numbat-1.12 -e '1+2'	34.2 ± 0.9	32.2	36.7	1.00

[sharkdp]

Thanks to various changes by @rben01 (#559, #569, #578, #579, #585, #597, #598 ), there is another 5% improvement in startup time. For more details, see #579 (comment) (note: the absolute times are not comparable. I measured on different machines).

Also, note that some of the measurements above might be underestimating the true performance benefit, since Numbats prelude is growing. For a true comparison, we need to run both versions against the same version of the prelude (typically the older one, as the new one might not be supported on the old interpreter).