Avoid excessive retypechecking of TH codebases

[pepeiborra]

This PR follows the plan outlined in #660 including tests.

Our current benchmark example (Cabal) doesn't include any example of TH, so the only thing visible in the benchmark results is that things do not regress too much:

version	name	success	samples	startup	setup	experiment	maxResidency
upstream	code actions	True	100	10.000944527000001	0.59697904	0.68091352	196MB
HEAD	code actions	True	100	9.981281715000001	0.9265053580000001	0.905302617	204MB
upstream	code actions after edit	True	100	9.870548535000001	0.0	29.209360251000003	197MB
HEAD	code actions after edit	True	100	10.058724974	0.0	31.147682330000002	205MB
upstream	completions after edit	True	100	10.079300106	0.0	38.186949398	173MB
HEAD	completions after edit	True	100	10.005696135	0.0	41.423665788	183MB
upstream	documentSymbols after edit	True	100	9.973052604000001	0.0	4.571069307	170MB
HEAD	documentSymbols after edit	True	100	10.009694254000001	0.0	4.464170808	170MB
upstream	edit	True	100	10.684809285	0.0	27.757133392	168MB
HEAD	edit	True	100	10.892325944000001	0.0	29.814292458	172MB
upstream	getDefinition	True	100	10.52725915	0.0	0.8384268300000001	172MB
HEAD	getDefinition	True	100	11.522601252000001	0.0	0.8449014170000001	171MB
upstream	hover	True	100	10.000964431	0.0	0.441070864	170MB
HEAD	hover	True	100	9.916935132	0.0	0.48269071300000005	171MB
upstream	hover after edit	True	100	10.222695908	0.0	35.27605837	170MB
HEAD	hover after edit	True	100	10.497113535	0.0	38.391144658	178MB

Full results in https://github.com/pepeiborra/ghcide/tree/TH-benchmarks/bench-hist

I also noticed that TH reloading is not happening when not using interface files, that is, when the TH is in a file of interest. Raised #672 and included a failing test here.

[cocreature]

Thanks!