Make CPU's IrEmitter::hlo_to_profile_idx_ a value

I think the performance advantages of keeping it as a maybe-null pointer are
minimal, and it instead complicates the signature generation code.  For example,
the code to generate calls to __xla_cpu_runtime_ParallelForkJoin is buggy when
hlo_to_profile_idx_ is nullptr today.  This bug isn't visible today because we
always have hlo_to_profile_idx_ as nullptr in JIT mode and in AOT mode we don't
parallelize Hlo operations.

PiperOrigin-RevId: 175993645