[raphael]

Generating code and running the generator test suites on large multi-service designs spent most of its time in repeated attribute deep copies, template re-parsing, full-file import formatting, reflection path allocations during design snapshot checks, and sequential go test -mod=mod subprocesses. This change reduces both goa gen execution time and the code generation test suite runtime while upgrading the module minimum to Go 1.27.3.

During code generation:

• codegen.mappedObjectDefinition previously called expr.NewMappedAttributeExpr, which recursively cloned entire nested attribute graphs (DupAtt) just to strip :tag suffixes from top-level object fields during transform planning. expr.NewShallowMappedAttributeExpr now copies only the top-level object attribute and its immediate field slice while preserving shared UserType origin identity across sibling fields.
• SectionTemplate.Write now caches parsed text/template instances for templates using the default TemplateFuncs() map instead of reparsing template text on every rendered section.
• finalizeGoSource now formats only the package and import header through the first non-import declaration when imports.Process (FormatOnly: true) runs on files without //line directives, splicing the formatted import block onto the already-printed AST body instead of re-parsing and re-printing the entire file a second time.
• captureDesignState now reuses a growable path buffer, caches struct field metadata per reflect.Type, and hashes directly into fnv.New128a() without allocating intermediate path strings on every visited node.

In the codegen test suites (codegen, codegen/service, codegen/generator, http/codegen, grpc/codegen, and jsonrpc/codegen), temporary test modules now pre-tidy go.mod and go.sum once per test binary and compile generated packages in parallel using -mod=readonly (and go build when no generated *_test.go files exist), cutting the combined test run from ~222s to ~56s.