fix(asm): encode the loong64 64-bit-span 2RI14 offsets like the toolchain

Assisted-by: GLM 5.3

asm/loong64_encode.go

@@ -264,9 +264,8 @@ func l64WordsLE(ws ...uint32) []byte {
 // a 32-bit one (C_LOREG_32) splits between addu16i.d (bits 31:16, materialised
 // in R30, the assembler temp) and si14 (bits 15:2); anything wider
 // (C_LOREG_64) builds the whole constant in R30 through lu12i.w + ori +
-// lu32i.d + lu52i.d and folds the base into it.  The loong64 backend resolves
-// memory offsets as int32, so the third span answers nothing the encoder can
-// ask today: it stands so the mapping stays complete if that domain widens.
+// lu32i.d + lu52i.d and folds the base into it.  The caller adjusts the
+// opcode for the 64-bit span's load-direction quirks before arriving here.
 func l64firr14Words(op uint32, off, rj, rd int) []byte {
 	switch {
 	case off >= -32766 && off < 32766: