Go: model for ... range as a rich Go.RangeLoop instead of cramming it into J.ForEachLoop

[knutwannheden]

Motivation

Running a recipe over Go source mangled every for ... range loop:

-	for _, v := range items {
+	for_:=range items {

Go's range loop binds a key and an optional value joined by := or = (for k, v := range x, for k := range x, for range x). That doesn't fit Java's single-variable J.ForEachLoop.Control (one variable + iterable). The Go side worked around it by defining its own fatter java.ForEachControl (Key/Value/Operator/Iterable) masquerading under the Java type name, then bridging it to the real J.ForEachLoop.Control over RPC with a hand-written codec.

• That bridge is the bug: the real J.ForEachLoop.Control has nowhere to put a second variable or the :=/=, so the codec synthesized a single variable from the key and dropped the value and the operator. Like the operator fixes in Go: map Go-specific operators to rich Go.* LST nodes so they survive RPC round-trips #7842, it round-trips fine Go↔Go (both sides use the fat struct) but truncates on the Java leg — invisible until a recipe forces full deserialization.

Examples

New node (mirrors the per-language loop-node pattern — JS has ForOfLoop/ForInLoop):

Go.RangeLoop   for k, v := range x { … }
  key       JRightPadded<Expression>   // nullable (for range x)
  value     JRightPadded<Expression>   // nullable (for k := range x)
  operator  JLeftPadded<Type>          // Type {Equals, Define}; before = space before `range`
  iterable  Expression
  body      JRightPadded<Statement>

Summary

• Added Go.RangeLoop end-to-end (Java Go.java/GolangVisitor/GolangSender/GolangReceiver; Go tree + parser + printer + visitor + RPC sender/receiver + factory & RegisterValueType).
• Removed the cheat entirely: the Go-customized java.ForEachLoop/java.ForEachControl structs and methods, their RPC send/receive, the lossy visitForEachControl/visitForEachLoop bridge overrides, and the J$ForEachLoop/J$ForEachLoop$Control wire registrations. Template matching (comparator/substitution/go_template) now matches Go.RangeLoop.
• AssignOp (Equals/Define) is kept and reused as the operator enum, round-tripping as the Go.RangeLoop.Type constant name.

Note for reviewers: a Go LST node needs both RegisterFactory (receive) and RegisterValueType (send) — registering only one silently desyncs the wire into a ParseError.

• This is one of a series removing "cheats" where a Go-side java.* struct carries more than its real J.* counterpart and truncates over RPC (follow-up to Go: map Go-specific operators to rich Go.* LST nodes so they survive RPC round-trips #7842). Known remaining, each its own follow-up PR: multi-value return, if init; cond, switch init; tag, grouped var(...)/const(...) blocks.

Test plan

• New range cases in GoRpcPrintRoundTripIntegTest (key+value, key-only, no-variable, = assign) — all green
• ./gradlew :rewrite-go:test :rewrite-go:integTest passes (no regressions)
• go test ./... passes