Use safer clangd defaults for enrichment

[darthghandi]

Summary

• Change the built-in clangd enrichment defaults to disable background indexing and clang-tidy, keep header insertion disabled, and limit clangd to one worker.
• Add a regression test that pins the default clangd argument list.
• Document the new defaults and how users can override them with semantic.providers[].args.

Test Plan

• GOCACHE=/tmp/gortex-go-cache GOMODCACHE=/tmp/gortex-gomod-cache go test ./internal/semantic/lsp -count=1
• Manual llama.cpp diagnostic run confirmed gortex launched clangd with --background-index=false --clang-tidy=false --header-insertion=never -j=1.
• go test ./... currently fails on both this branch and main in internal/parser/forest due to github.com/alexaandru/go-sitter-forest/perl C compilation error: bsearch.h:72:7: error: expected identifier or ‘(’ before ‘_Generic’.

Related: #220

[zzet]

@darthghandi --background-index=false has caveats:

It can impact the C/C++ cross-file graph fidelity:

• find_usages / get_callers on a C/C++ function -> only same-TU callers; call sites in other .cpp files are not found. Ambiguous call edges never get confirmed to the lsp_resolved tier, so min_tier filtering drops them → missing caller edges.
• find_implementations -> misses implementers in other files.
• Type hierarchy -> misses subtypes defined elsewhere.

I want to benchmark it before accepting it.