[bentsherman]

The engine rewrite (#194) moved parsing onto the engine thread, because nf-lang writes to Groovy's shared ClassHelper nodes when it parses a fully qualified type such as java.lang.String. That made initial indexing slower.

That write only happens for fully qualified names of a few core types, which are rare in practice and discouraged, and it is being fixed in nf-lang. This PR restores parallel parsing: each read-pool task now reads the file and then parses it into its own SourceUnit. Registering the unit with the compiler, include discovery, and all analysis stay on the engine thread.

• Builder.parseAll parses in the read-pool task. A parse exception is reported as an internal error for that file, as before.
• Docs: the thread table and invariants in docs/internals.md, and the "Work off the engine thread" section of the core ADR.

Benchmark

make bench-compare against main on nf-core/rnaseq (3 alternating rounds, 20 iterations per scenario):

scenario	main	this PR	delta	verdict
workspace indexed	4763 ms	3925 ms	-17.6%	faster (p < 0.001)
startup: initialize	426 ms	414 ms	-2.7%	noise
didOpen → diagnostics	503 ms	504 ms	+0.2%	noise
didChange → diagnostics	791 ms	787 ms	-0.4%	noise
peak RSS	708 MB	823 MB	+16.2%	noise (p = 0.08)

Requests (hover, completion, references, etc.) are unaffected. Peak memory is likely higher because several files are parsed at once during indexing, but it isn't significant over 3 rounds.