Fix backward jump optimization in generic for loops

[sjh9714]

Fixes #540.

Changes proposed in this pull request:

• Stop following jump chains when they reach the current or an earlier instruction. Earlier jumps have already been patched from label IDs to relative offsets; interpreting those offsets as labels can make a generic for with a constant-false branch lose return values, panic, or loop forever. Leaving the backward jump intact preserves its target without changing the VM.
• Add regression cases for the reported return, sum, and preceding-loop failures, dead branches containing break or a nested loop, and runtime-false/numeric-loop controls. Five cases fail before the fix; all seven pass afterward.

Testing passed:

• go test -count=1 ./...
• go test -race -run '^TestGenericForConstantFalseBranch$' -count=1 -v .
• make build
• go build ./...

Additional checks found existing baseline failures: go vet ./... reports self-assignment and unreachable code, and go test -race -count=1 ./... reports a finalizer race in TestLocalVarFree. The vet diagnostics are identical on pristine upstream; go test -race -run '^TestLocalVarFree$' -count=1 . also reproduces the race there.