Support scalable offset in CFI.
Currently, targets with scalable vectors (AArch64, RISCV) handle
scalable offsets in cfi using cfi_escape expression. The problem is that
CFIInstrInserter doesn't understand these expressions.
We define new "pseudo" cfi instructions to support scalable offsets in
CFI:
(1) llvm_def_cfa_reg_scalable_offset - creates the new rule for calculating
cfa: `deref(Reg + ScalableOffset * x + FixedOffset)` where `x` is the
"scale" runtime constant.
(2) llvm_reg_at_scalable_offset_from_cfa - creates the new rule to calculate
previoius value of register `deref(cfa + ScalableOffset * x +
FixedOffset)` where `x` is the "scale" runtime constant.
(3) llvm_reg_at_scalable_offset_from_reg - creates the new rule to calculate previous value
of register `deref(Reg2 + ScalableOffset * x +
FixedOffset)`. This rule is to be used when
offset from cfa is not known, but intead fixed offset from `Reg2` is
known.
All of these cfi_instructions will be "lowered" to escape expressions
during final assembly emission. But during `CFIInstrInserter` these are
not lowered yet, so their semantics can be understood without decoding
cfi expressions. Since (1) and (2) depend on how target calculates
scalable offsets, their lowering is target-dependent.
